# Zero to Agent: Vercel x Deepmind Hackathon NYC: Project Gallery

- **Event:** [Zero to Agent: Vercel x Deepmind Hackathon NYC](https://cerebralvalley.ai/e/zero-to-agent-nyc)
- **When:** Sat, Mar 21 at 9:00 AM – 10:00 PM (EDT)
- **Where:** 220 W 26th St, New York, NY
- **Hosts:** [Vercel](https://cerebralvalley.ai/u/vercel), [Google DeepMind](https://cerebralvalley.ai/u/deepmind)
- **Projects:** 62 (6 placed)
- **Page:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery

## Projects

### 1. Watermelon Sugar

DevOps teams waste hours manually checking dashboards, triaging incidents, and writing fix PRs across multiple projects. AI DevOps Agent is an autonomous agent that monitors your Vercel infrastructure 24/7 — detecting deployment failures, running uptime checks, scanning for security misconfigurations, and automatically opening GitHub PRs with fixes. Teams interact with it via a chat UI or directly from Slack using /devops <request>. It runs on a schedule (hourly monitoring, weekly digest) and takes real action — not just advice.

- **Team:** [Ronit Gandhi](https://cerebralvalley.ai/u/RonitGandhi), [Vanshika Ram Gurbani](https://cerebralvalley.ai/u/VanshikaGurbani)
- **GitHub:** https://github.com/sugarwatermelon300-crypto/zerotoagent
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=58

### 2. Parallax

Parallax is an autonomous UX testing agent that launches 5 AI personas simultaneously to navigate any website and report real friction. Powered by Gemini 2.5 Computer Use API, each persona actually sees and interacts with your site like a real user — clicking, scrolling, reading. Speedrun Steve finds unnecessary friction. Confused Clara spots unclear labels. Skeptical Sam flags dark patterns. Accessible Alex catches a11y failures. Global Gita finds i18n gaps. Results stream live with screenshots, thought bubbles, and a Parallax Score in under 60 seconds. No static analysis — real agent journeys on real browsers.

- **Team:** [Kanishkha Jaisankar](https://cerebralvalley.ai/u/Kanishkha_J)
- **GitHub:** https://github.com/Jkanishkha0305/parallax
- **Demo video:** https://vimeo.com/1175813183?share=copy&fl=sv&fe=ci
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=59

### 3. Parent Teacher Bridge

This app enables teachers to communicate with parents in a way that leverages the best features of AI to organize and elaborate feedback asynch. A parent can also engage the teacher in real time communication. The parent is then able to ask the agent to set time on their calendar based on their time availability to help their student.

- **Team:** [Marlon Fuentes](https://cerebralvalley.ai/u/MarlonFuentes)
- **GitHub:** https://github.com/MarlonCreative/parent-teacher-bridge
- **Demo video:** https://youtu.be/mjxRNdlsjFg
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=60

### 4. Team Titans

StyleAgent is a voice-first personal shopping agent that builds complete, shoppable outfits tailored to your style, occasions, and budget in real time. Powered by Gemini's multimodal reasoning, StyleAgent connects directly to catalogs across all major fashion brands, giving users one unified destination to discover, curate, and shop.The agent doesn't just recommend products, it learns. By continuously analyzing each user's interactions, browsing behavior, and styling sessions, StyleAgent builds a deeply personalized shopping experience that evolves with them over time. Every session with your AI stylist makes the next one smarter.No more switching between apps, no more generic recommendations. StyleAgent is your personal stylist, your wardrobe curator, and your smartest shopping companion all in one place.

- **Team:** [saumay killa](https://cerebralvalley.ai/u/saumaydev), [Folusho adeyemi](https://cerebralvalley.ai/u/folusho), [Jeremiah Richard](https://cerebralvalley.ai/u/drjay)
- **GitHub:** https://github.com/saumaykilla/AI-Hackathron-StyleAI
- **Demo video:** https://drive.google.com/file/d/1QrpZSJNDMuunkk9TUFx01CMQwfiSdzpI/view?usp=drive_link
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=61

### 5. Structured

automating from unstructured 3D models to structured 3D models that is engineer-grade and manufacturable

- **Team:** [Dien Hu](https://cerebralvalley.ai/u/de_an_hu)
- **GitHub:** https://github.com/deanhu0822/Structured
- **Demo video:** https://youtu.be/Enp0JcrP--I
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=62

### 6. DOMinion

Online articles are full of unverified claims, missing context, and jargon that goes unexplained. Readers have to fact-check things themselves, open a dozen tabs, and still walk away unsure what's true.

DOMinion is a Chrome extension that brings something like Community Notes to every article you read, except it's instant and powered by AI research instead of waiting for crowdsourced votes. Click the extension, and it scans the article and annotates it with four types of highlights:

- **Fact Check**: verifies claims against other sources the agent finds through deep research
- **Missing Context**: surfaces important background or framing the article left out
- **Explanation**: breaks down jargon, technical language, or complex concepts
- **Source Analysis**: follows the article's cited sources and evaluates whether they actually support the claims being made

Each annotation pulls from both the article itself and from other articles the agent researches in real time. Hover any highlight to see a summary. Click it to start a conversation and dig deeper. It works on news, opinion pieces, and arXiv papers.

**Built with:** Gemini LLMs (3.1 Flash Lite for speed or Pro for thoroughness depending on the article's complexity), Google Search grounding, Chrome Extensions API.

- **Team:** [Nathan Cooper](https://cerebralvalley.ai/u/natedawg), [Danil Meresenschi](https://cerebralvalley.ai/u/mereshd)
- **GitHub:** https://github.com/mereshd/DOMinion
- **Demo video:** https://youtu.be/eSXUzZdpvdc
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=1

### 7. Secret Agent

Patchwork is a GitHub agent built at the Zero to Agent: Vercel x Google DeepMind Hackathon in NYC.

File a bug report as a GitHub issue. Patchwork automatically reads your codebase, writes a regression test using Gemini 2.5 Pro, and opens a PR — so that bug can never silently ship again.

🔗 Live demo: https://pensive-jang.vercel.app
🔗 GitHub: https://github.com/rohan1402/patchwork
🔗 Demo repo: https://github.com/rohan1402/patchwork-demo

⚙️ Tech Stack
- Gemini 2.5 Pro via Vercel AI SDK — reasoning about bug reports and generating tests
- Next.js on Vercel — webhook receiver and dashboard
- GitHub API via Octokit — reading codebases, creating branches and PRs
- Supabase — logging runs for the live dashboard
- waitUntil from @vercel/functions — background processing after webhook response

Built by Rohan Pant
🔗 linkedin.com/in/rohan1402

- **Team:** [Rohan Pant](https://cerebralvalley.ai/u/rp1610)
- **GitHub:** https://github.com/rohan1402/patchwork
- **Demo video:** https://youtu.be/pvlw-jDOaSo
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=2

### 8. Iron Panda

Here's the full README:

---

# Iron Pando Dome 🐼🛡️

**Your Guardian AI — Protecting You From Scams, One Conversation at a Time**

Iron Pando Dome is a mobile-first web application designed to help people — especially seniors and vulnerable populations — identify and avoid online scams. Powered by an AI assistant named **Pando**, the app provides real-time, conversational guidance through voice and text.

## What It Does

### 🎙️ Talk to Pando (Voice & Text Chat)
- **Hands-free voice loop**: Tap the mic, speak, and Pando auto-detects when you're done (2-second silence detection), responds with voice, then listens again
- **Streaming AI responses** and **Streaming TTS** for ultra-low latency
- **Multiple voice options**

### 📧 Email Scanner
Paste a suspicious email and Pando analyzes it for phishing, social engineering, and red flags.

### 📚 Scam Education Library
Curated knowledge base of common scam types and safety tips.

## Tech Stack

**Frontend**: React 18, TypeScript, Tailwind CSS, Vite, React Router, TanStack React Query, Radix UI

**Backend**: Supabase Edge Functions (Deno) — `pando-chat`, `elevenlabs-tts`, `elevenlabs-stt`, `scan-email`

**AI & Models**: Lovable AI Gateway → Google Gemini 2.5 Flash (chat, email analysis, vision, tool-calling)

**Voice**: ElevenLabs TTS (streaming, `eleven_turbo_v2_5`) + STT (Scribe `scribe_v2`), Web Audio API silence detection, MediaSource API streaming playback

**Hosting**: Lovable Cloud + Supabase

## Team

**Iron Panda** team: **Forrest Pan** & **Frank Yu** Ji Thakur**Wun Kuen Ng

---

*Built with ❤️ to make the internet a safer place.*

- **Team:** [frank yu](https://cerebralvalley.ai/u/frankyu)
- **GitHub:** https://github.com/Mkultra00/ironpandadome
- **Demo video:** https://youtu.be/JTLDUR6mryc
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=3

### 9. Prabhu Kannan_NYC_Hackathon_3_21_26

AgentPact turns plain-English business policies into battle-tested, machine-enforceable AI agent governance — in 30 seconds instead of 6 weeks.The Problem: Every company deploying AI agents needs governance, but today it takes security teams weeks to manually write rules, engineers months to build enforcement layers, and the result is static — unable to handle adversarial behavior or evolve.

The Solution: A 5-agent adversarial pipeline powered by Google Gemini 2.5 Flash.

1. Charter Generator — converts natural language policy into structured governance rules
2. Red Team Adversary — attacks the charter to find loopholes, split attacks, and exploits
3. Resolver — automatically patches every vulnerability found
4. Enforcement Engine — evaluates runtime agent decisions against the hardened charter
5. Amendment Agent — proposes charter evolution from observed violations

AgentPact ships with 15 industry-specific policy packs (Financial Services, Healthcare/HIPAA, Autonomous AI Agents, Legal, AdTech, Government, Education, Cybersecurity, Manufacturing, Real Estate, Energy, E-commerce, HR, Biotech, Travel) and includes a compliance scoring system, runtime enforcement dashboard, governance diff view, and Supabase persistence.

Built with Next.js 15, deployed on Vercel, routed through Vercel AI Gateway with automatic fallback, authenticated with Clerk, monitored by Sentry, persisted via Supabase, and audio briefings via ElevenLabs.

- **Team:** [Prabhu Kannan](https://cerebralvalley.ai/u/prabhukannan)
- **GitHub:** https://github.com/pkanna000/agentpact
- **Demo video:** https://youtu.be/H22iQQqqIEM
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=4

### 10. Valiron

A2A Trust Gate is an AI agent orchestrator that solves the trust problem in agent-to-agent commerce. When one AI agent needs to delegate a high-stakes task — like purchasing a product — to another agent, there's currently no way to verify whether that agent is trustworthy or efficient before handing over control.

A2A Trust Gate fixes this by using Valiron's on-chain reputation layer (ERC-8004) to gate every delegation in real-time. Each agent has a verifiable trust score, a Moody's-style credit tier (AAA through C), and a routing decision (production, sandbox, or blocked). Only agents above a configurable trust threshold are allowed to execute.

In our live demo, three checkout agents with identical skills but wildly different trust profiles compete for a real purchase. The trusted agent (score 93, AAA) passes the gate and places a real order via the Zinc API. The unverified agent (score 0) is automatically blocked. No mocks, no simulations — real on-chain data, real orders.

Built with Next.js on Vercel, Vercel AI SDK, Google Gemini 2.5 Flash, and Google's A2A protocol (agent card at /.well-known/agent.json). The UI was scaffolded with V0 and features a Command Center with streaming tool invocations and an Intel Panel showing live agent trust data.

- **Team:** [Drew Mailen](https://cerebralvalley.ai/u/dr3w), [Vatsa Shah](https://cerebralvalley.ai/u/0xVatsaShah)
- **GitHub:** https://github.com/vatsashah45/A2A-Trust-Gate
- **Demo video:** https://vimeo.com/1175805992
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=5

### 11. focusgroup.ai

https://focusgroup-ai.vercel.app/

focusgroup.ai is an AI-powered reactive marketing platform that helps consumer brands turn live cultural moments into launch-ready campaigns in minutes.

Brands constantly miss timely opportunities because reacting to breaking sports, entertainment, or culture moments usually requires too many manual steps: finding the story, drafting creative, choosing the right product, reviewing for audience fit, checking risk, building channel-ready content, and deciding who to send it to. By the time that process is done, the moment is gone.

focusgroup.ai solves that by compressing the entire workflow into one system. A brand can paste a live news story or select a trending headline, choose the brand line they want to promote, and instantly generate multiple campaign directions tied to the actual story context. The system then runs a built-in AI focus group with distinct audience segments, scores the concept, rewrites the draft from feedback, generates creative assets, previews the final campaign as a branded email, auto-selects audience segments, and sends the campaign.

This turns reactive marketing from a slow creative bottleneck into a fast testing loop. Instead of guessing what might work, brands can validate message quality, product fit, and audience response before committing larger budget.

What It Does

Pulls in live or fallback news/trend stories
Reads the article context and generates campaign concepts
Matches the campaign to a real product catalog
Runs AI audience review with segmented focus-group personas
Rewrites content based on feedback to improve performance
Generates branded creative and email-ready campaign output
Selects likely audience segments for launch
Simulates measurable outcomes like opens, clicks, and conversions

What Problem It Solves
Modern brands need to move at internet speed, but their workflow does not. Creative teams, marketers, legal reviewers, and channel owners all slow down the process. focusgroup.ai gives brands a way to react quickly without losing message quality, product relevance, or audience targeting. It is especially useful for consumer brands in beauty, sports, fashion, snacks, and entertainment-adjacent categories where timing directly affects campaign performance.

- **Team:** [Henry He](https://cerebralvalley.ai/u/handehehehe)
- **GitHub:** https://github.com/Henryhehe/focusgroup.ai/tree/main
- **Demo video:** https://youtu.be/kkzar54FejE
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=6

### 12. Swwagger

human in the loop to take AI slop to ai masterpiece

- **Team:** [Ethan Huang](https://cerebralvalley.ai/u/3than)
- **GitHub:** https://github.com/3thanh/GenReview
- **Demo video:** https://www.loom.com/share/ea823ddf33bc4a17bbb6574326a9935b
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=7

### 13. POView

POView redefines digital exploration by turning the world into an interactive conversation. Leveraging a state-of-the-art 3D geospatial engine and the Gemini Live API, it lets you soar over cities in cinematic flyovers while speaking naturally with an intelligent assistant to uncover deep, real-time insights about the neighborhoods below.

- **Team:** [Divija Naredla](https://cerebralvalley.ai/u/Divicodes), [Shaurya tiwari](https://cerebralvalley.ai/u/shazam)
- **GitHub:** https://github.com/Shazam6565/POView-v1.0.2
- **Demo video:** https://www.youtube.com/watch?v=Z5yv-IPWc-g
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=8

### 14. Flock

Flock is an AI-powered group travel planner that replaces the broken group chat → poll → someone-just-books-it cycle. Each person gets a personal AI agent that learns their travel taste from text, voice, and photos — not just "beach or city" but  textures, lighting, energy, budget tolerance, and aesthetic fingerprint via Gemini Vision embeddings. Agents then negotiate on behalf of each traveler in a live 3-round session: constraints, vision, and dealmakers. A mediator agent resolves  conflicts using pgvector similarity matching against 194 pre-embedded hotels across 8 cities. The result: a personalized trip that genuinely satisfies everyone, revealed with a cinematic split-screen showing the negotiation timeline and each person's personalized card — including a "surprise" recommendation nobody asked for, derived purely from their photos. Built in 7 hours at Zero to Agent.

- **Team:** [Rahil Singhi](https://cerebralvalley.ai/u/rahil), [Aishwarya Ghaiwat](https://cerebralvalley.ai/u/aishghaiwat)
- **GitHub:** https://github.com/rahilsinghi/Flock
- **Demo video:** https://youtu.be/MZmUSAgNpt4
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=9

### 15. midtown-money

Pick your nightmare client — the skeptical investor who already has an advisor, the busy professional who won't give you five minutes, the retiree who's afraid of losing everything — and have a live AI voice conversation that pushes back exactly like they would. When the call ends, Midtown Money shows you the exact line where you lost them, the exact words you should have said instead, and gives your manager a live view of who on the team is improving and who needs help.

- **Team:** [Shubhan Kadam](https://cerebralvalley.ai/u/shubhvn), [Pranav Raj Narayanan Kavitha](https://cerebralvalley.ai/u/kimii), [Sumanth Ramesh](https://cerebralvalley.ai/u/sumanthssr), [Saketh Raman Ramesh](https://cerebralvalley.ai/u/sakethraman)
- **GitHub:** https://github.com/Sumo-99/midtownmoney
- **Demo video:** https://www.youtube.com/watch?v=oPlayP5y0Lk
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=10

### 16. MIRA

The AI agent uses acuppressure points to assist users to relieve pain points like anxiety, nervousness, fear, etc... AI agent will ask for user feedback and recommend another cycle or other points.  The AI agent starts with breathing techniques to calm the vagus nerve and then it  ask for user feedback and then make other recommendations for users until users said, "I m done."

- **Team:** [KOA Wunkuen Ng](https://cerebralvalley.ai/u/WKN)
- **GitHub:** https://github.com/wunkuen/soothe-spirit-bot
- **Demo video:** https://youtube.com/shorts/GISxWilHl6E?feature=share
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=11

### 17. Pre-Life

PreLife is an AI-powered life decision game where players explore "what-if" scenarios by walking through a pixel-art city and making binary choices that branch into different futures. When a player enters a scenario like "Become a Developer," 8 specialized AI agents (powered by Gemini) simultaneously research real-time industry trends, job market data, social media sentiment, risk factors, and success strategies, displaying their findings live as the city loads. Players then navigate an interactive pixel-art world, encountering NPCs who present meaningful binary choices (Cornell vs NYU, Research vs Startups, PhD vs Industry). Each choice affects a life bar, generates AI-illustrated social media posts, and unlocks new paths with new characters. At the end, a detailed outcome report compares the player's chosen path against the alternative with percentage-based scoring, key factors analysis, and personalized advice. PreLife solves the problem of decision paralysis. Instead of static career advice or generic personality quizzes, it lets users experience the consequences of their choices through an engaging, gamified experience grounded in real-world data.

- **Placement:** Finalist
- **Team:** [Chinmay Shringi](https://cerebralvalley.ai/u/asdf), [Sariya Rizwan](https://cerebralvalley.ai/u/sariyarizwan)
- **GitHub:** https://github.com/ChinmayShringi/game_of_life
- **Demo video:** https://youtu.be/_abn3Xt2XSE
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=12

### 18. MORPH

Morph turns one sentence of natural language into a running, device-aware tool — composed in real time from capability blocks, planned by Gemini 2.5 Flash, with AI-generated custom React UI per block, validated in Vercel Sandbox, and rendered live with permission gating and mid-session mutation.

You type or speak an intent. Gemini reasons over a library of 6 capability blocks (Timer, Step Counter, Voice Recorder, Compass, Camera + Vision, Map) and your device's actual sensor availability, then generates a ToolSpec. Each block gets a purpose-built React component generated by Gemini — not a generic template, but code tuned to your exact words. That code runs in an isolated Vercel Sandbox microVM before it ever reaches your browser. Only validated code renders, with a "Sandbox-verified" badge.

Same intent, different tools: a laptop gets timer + camera + voice, a phone gets steps + compass + map. Mid-session you can say "add a rating button" — new block appears while your timer keeps counting and your steps keep tracking. Mutations are additive only, state never resets.

The block library is also exposed as an MCP server, so any MCP-compatible agent (including Claude) can discover and compose Morph tools programmatically. Morph is the runtime. Intent in, tool out.

- **Team:** [VISHAL Sunil Kumar](https://cerebralvalley.ai/u/Vshal), [Sanika Sawal](https://cerebralvalley.ai/u/sanikasawal)
- **GitHub:** https://github.com/sanikasawal/Morph
- **Demo video:** https://youtu.be/RZJOK8y3tPU
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=15

### 19. hive

Hive is designed for the way people actually use AI in code today, where you have three terminals open, you're running gemini on one feature, your friend is running it on another, and in five minutes you're already in a merge conflict you didn't see coming because neither of these agents knows what the other is doing. That's not an enterprise problem, that's a vibe coder problem. You're moving quickly, you're shipping at 2 am, and your entire workflow is you and a few friends just vibing on a code base, throwing prompts at it, and you don't want to have to deal with PRs and code reviews, you want to not have to step on each other. Hive integrates directly into the gemini-cli you're already using, it doesn't require any config changes, and it gives all the agents in your session a live view of what the others are doing so you naturally stay out of each other's way.

- **Team:** [Shresth Kapoor](https://cerebralvalley.ai/u/shresthkapoor7), [Aditya Maheshwari](https://cerebralvalley.ai/u/Adi_on_caffeine)
- **GitHub:** https://github.com/shresthkapoor7/hive3
- **Demo video:** https://www.youtube.com/watch?v=A5Mq5H0uHX8
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=13

### 20. FixFlow

Live App Link: https://fix-flow-seven.vercel.app

FixFlow is an AI-powered property maintenance agent that turns a single photo into a fully dispatched repair. Diagnosis, contractor discovery, vetting, quoting, and approval are all automated.

Here's the problem: 10 million independent landlords in the US coordinate maintenance through text messages, phone tag, and Google searches. Reactive repairs cost 3-5x more than preventive ones, and slow response drives tenant turnover that costs $4,000 per unit on average.

Here's how FixFlow solves it: A tenant photographs a broken sink. Within seconds, Gemini Flash's vision capability diagnoses the issue (plumbing emergency, severity 5/5), flags a safety warning, and recommends immediate action. The agent then uses Gemini's Google Maps tools to find real licensed plumbers within 3 miles of the property. These are real businesses with real ratings, hours, and phone numbers pulled live from Google Maps. Each contractor is then vetted using Gemini's Search grounding: the AI summarizes real reviews, flags licensing concerns and red flags, estimates repair costs ($163 to $2,040), and links every finding to its source.

The tenant selects a contractor, and FixFlow contacts them automatically via an AI-powered phone call using ElevenLabs voice generation delivered through Twilio. The agent explains the issue, provides the details, and asks the contractor for their quote. The contractor states their price over the phone, and the AI extracts it in real time. The quote then appears live on the tenant's dashboard in the Contractor Quote section, no manual entry needed. If the contractor doesn't answer, a detailed email is sent as a fallback with the full issue details, photo evidence, and cost estimate. The entire background pipeline of finding, vetting, and notifying contractors is orchestrated through Vercel Workflows, making the process durable and retryable. An AI voice briefing also keeps the tenant informed at every step of the process.

Meanwhile, the landlord has a real-time Command Center showing all tickets across all properties: 20 active requests, 4 needing approval, $42K monthly spend, average resolution under 1 hour, with one-click approval for dispatches.

One photo. Full dispatch.

- **Team:** [Varun Bhandari](https://cerebralvalley.ai/u/vbhandari), [Vraj Chetankumar Patel](https://cerebralvalley.ai/u/Vraj_19), [Chaitanya Manik](https://cerebralvalley.ai/u/Chaitanya17), [Ethan Choukroun](https://cerebralvalley.ai/u/echouk)
- **GitHub:** https://github.com/goffycoder/FixFlow.git
- **Demo video:** https://www.youtube.com/watch?v=9qYMq9ejLEg
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=17

### 21. Just Us and PM Steve

PM Agent Steve is an AI-powered project management assistant built to solve one of the most persistent pain points in software teams — the overwhelming amount of time PMs spend just to get a single ticket created. Between back-and-forth communications with developers, manual workload checks, effort estimations, and trying to understand unfamiliar tech stacks, ticket creation becomes a process that can drag on for days before any real work begins. Steve is not here to replace the PM. He is here to make the PM dramatically more effective. At every stage of the planning process, Steve presents his analysis and waits for human review and approval before moving forward — the PM stays the decision maker throughout. Steve handles the research, the breakdowns, the estimations, and the assignments, so the PM can focus on judgment rather than legwork. Once work begins, Steve continues to assist through a dynamic Mission Control dashboard that gives PMs real-time visibility into team velocity, bug impact, and scope drift across all active projects — making it easy to understand what the team is capable of, where things are slipping, and why. Built on Gemini 3 Pro with Google Search grounding, Supabase, and Linear, PM Agent Steve is designed around one belief: the best outcomes happen when human expertise and AI capability work together, not when one tries to replace the other.

- **Team:** [Kerr Tan](https://cerebralvalley.ai/u/kerrk07), [Jingzhen Sha](https://cerebralvalley.ai/u/Pompomina), [Tracy Zhang](https://cerebralvalley.ai/u/_tracyyyyy)
- **GitHub:** https://github.com/Ceiceiceii/ur_pm_agent_steve
- **Demo video:** https://youtu.be/SQ-m1nungnM
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=14

### 22. Water Pistol

Tinder style RPG where you gain points by keeping interests in the friend zone, building up a roster over time, so they give you more items trying to gain your interest in return.

- **Team:** [Dhruv Patel](https://cerebralvalley.ai/u/dhruv020699), [Devin Dawson](https://cerebralvalley.ai/u/unclemusclez)
- **GitHub:** https://github.com/dhruv0206/friendzone
- **Demo video:** https://youtu.be/j7YI5RtQkI0
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=23

### 23. Rahul and Yu Xiang

Project translates screenshot to code fix.

- **Team:** [Rahul Rajaram](https://cerebralvalley.ai/u/rahulrajaram), [Yu Xiang Zhang](https://cerebralvalley.ai/u/yuxz)
- **GitHub:** https://github.com/rahulrajaram/relay
- **Demo video:** https://youtu.be/XEm5S1vjXLw
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=37

### 24. Team Dez

Tessera is an AI-powered knowledge graph that turns scattered research into a connected visual map. The problem: when you're reading articles, watching YouTube videos, scrolling
  through tweets, and tracking markets — your insights are siloed across dozens of tabs with no way to see how everything connects. Tessera fixes that. You feed it any content — URLs,
  screenshots, text, live market feeds — and Gemini AI extracts the key entities and relationships, automatically building an interactive graph that reveals connections you'd otherwise
  miss. The more you feed it, the smarter it gets

- **Team:** [Dez Calimese](https://cerebralvalley.ai/u/dezcalimese)
- **GitHub:** https://github.com/dezcalimese/tessera
- **Demo video:** https://youtu.be/6Zg7mi16qO4
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=44

### 25. Zero to Competitive intelligence

Zero to Competitive intelligence for start up founder research and build and deploy apps

- **Team:** [Forrest Pan](https://cerebralvalley.ai/u/fpreactrn)
- **GitHub:** https://github.com/panforrest/Zero-to-Competitive-Intelligence
- **Demo video:** https://youtu.be/0ZVnvTmOcCU
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=47

### 26. Grimoire

Grimoire clones any writing voice. Paste text, get a statistical fingerprint (9 metrics: syllable density, sentence rhythm, punctuation cadence, lexical diversity). See it as an interactive 3D shape. Then watch a multi-round agent rewrite your prose toward any target author's profile.

The agent generates 3 candidates per round at varied temperatures, scores each against the fingerprint, keeps the best, and feeds the gap back. Over 5 rounds the output converges toward the target. A blind LLM judge validates whether the result passes as the target author. Ships with 6 pre-computed voices (Hemingway, Poe, Twain, Kafka, Woolf, Fitzgerald). All metrics compute client-side, all AI streams in real time.

- **Placement:** Finalist
- **Team:** [Alex Wu](https://cerebralvalley.ai/u/amadeusWoo)
- **GitHub:** https://github.com/TheApexWu/grimoire
- **Demo video:** https://youtu.be/2ck-DhFizjo
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=16

### 27. Sequencing Machines

An agent built on Vercel AI SDK to parse sequencing protocol and extract important details for downstream anlaysis, and a slack chatbot with chat SDK as well.

- **Team:** [Sequencing Machines](https://cerebralvalley.ai/u/seqmachines)
- **GitHub:** https://github.com/seqmachines/cdna
- **Demo video:** https://youtu.be/VIx5P-XdlrQ?si=vGQks_OyiHvB9u-_
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=19

### 28. Lastline

As long-running coding agents autonoumously ship larger and larger pull requests, the bottleneck shifts from writing code to reviewing it. These agents make product decisions across entire user flows, not just single screens. They decide how a modal behaves, how a page looks in light mode, and how users move through the product.

Today, reviewing that work is still painfully manual. A human has to open the preview, walk through the app step by step, compare what they see against what they expected, capture screenshots, and file bugs.

We built Lastline, an QA agent for AI-generated pull-requests.

When a pull request hits a repo, Lastline is triggered by a GitHub webhook. It creates a QA plan, runs the product with Playwright, records the important user flows, and stitches them into one review video.

That video is sent to the human reviewer on Telegram. The reviewer can leave simple timestamped feedback, and Lastline turns that into structured GitHub issues with screenshots, timestamps, and PR context.

- **Team:** [Sparsh Paliwal](https://cerebralvalley.ai/u/discoman), [Prashant Soni](https://cerebralvalley.ai/u/mrsoni)
- **GitHub:** https://github.com/prashantsonibps/lastline/tree/main
- **Demo video:** https://youtu.be/850nGpBn2kc
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=18

### 29. memopal

memopal is a memory layer for autonomous agents. drop any data export (spotify, claude conversations, google takeout, photos, text files) and a hierarchical swarm of gemini agents extracts entities, relationships, claims, and events into a 15 table semantic knowledge graph. the core architecture forces genuine multi agent delegation: the orchestrator has zero write tools and must delegate to worker and leaf sub agents, creating n depth agent trees. every extracted fact links back to source evidence. the output is a rest api that any autonomous agent can query for structured personal context. built with ai sdk v6, gemini 3. pro/flash/flash lite across three model tiers, next.js 16, sqlite, and a realtime agent canvas that streams every tool call via sse.

- **Team:** [Joshua Lin](https://cerebralvalley.ai/u/qtzx06)
- **GitHub:** https://github.com/qtzx06/memopal
- **Demo video:** https://youtu.be/guaBNJ7tMTU
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=20

### 30. JinTeam

Real-time AI Auditing for the Modern Web.
Sentinel UX is an autonomous design partner that watches your screen as you build, instantly flagging accessibility, security, and performance issues before they reach production.

The Problem
Manual audits are slow, expensive, and reactive. Developers often ship code with WCAG violations or security vulnerabilities because they lack real-time feedback during the creative process.

The Solution
A live, multimodal auditor powered by Gemini 3 Flash. It analyzes visual state and code context simultaneously, providing a 6-pillar diagnostic score and actionable fixes in seconds.

Tech Stack
Intelligence
Gemini 3 Flash + LangChain
Voice Synthesis
ElevenLabs API
Frontend
React + Tailwind + Framer
Data Viz
Recharts
Backend
Express + Resend + Supabase
Deployment
Vercel Edge Functions
System Architecture & Flow
1. Capture
Supabase Edge Proxy + HTML Stream
2. Analyze
gemini-3-flash-preview
3. Diagnose
6-Pillar Audit Scoring
4. Alert
ElevenLabs Voice Feedback
5. Report
Resend Stakeholder Email
Key Innovation Pillars
Live Vision Audit
Real-time screen capture analysis using Gemini-3-Flash-Preview.

Mobile Screenshot Support
Upload mobile screenshots for instant accessibility and UX analysis.

Multimodal Consultation
Ask design questions via voice or text while Gemini analyzes your live screen.

Interactive Feedback Loop
Select specific audit findings to challenge or discuss with Gemini in real-time.

6-Axis Diagnostics
Quantified scores for WCAG, Security, UX, Mobile, Spam, and Perf.

Voice Feedback
Natural language audio alerts via ElevenLabs for hands-free auditing.

Automated Reporting
Instant PDF-style email summaries sent to stakeholders via Resend.

Context Awareness
Combines visual screenshots with raw HTML source fetched via Supabase Edge Proxy.

Accessibility First
Built-in WCAG 2.1 compliance checking as a core engine feature.

Cloud Persistence
Secure audit history and user profiles powered by Supabase and Vercel.

Market Impact
Democratizing Web Quality.
By providing enterprise-grade auditing tools for free to every developer, we can ensure the next generation of the web is more accessible, secure, and user-friendly by de

- **Team:** [Jitender Thakur](https://cerebralvalley.ai/u/jinreal)
- **GitHub:** https://github.com/jinthakur/VercelXdeepmind
- **Demo video:** https://youtu.be/XgZhIVRzEX8
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=21

### 31. Inspira

Our system helps designers and builders move from scattered inspiration to context-aware creation.

It works in two layers:

Layer one: design memory
We store references, patterns, and inspiration in a structured repository so valuable design context does not get lost.

Layer two: generation with taste
That stored context is then used by a coding agent to create more thoughtful design deliverables, grounded in actual inspiration and design intent rather than default patterns.

So instead of generating another polished but predictable interface, the agent produces outputs that feel more curated, relevant, and creatively informed.

- **Team:** [HUKHTA PATEL](https://cerebralvalley.ai/u/Hukhta23), [Jian Jin Chen](https://cerebralvalley.ai/u/JJC)
- **GitHub:** https://github.com/HukhtaPatel/v0-inspiration-library-app
- **Demo video:** https://www.loom.com/share/70c5948dc46b40729008970391f739c1
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=22

### 32. nan

GenUI Test Harness
AI-generated UIs look great in screenshots. But do they actually work? GenUI Test Harness is an autonomous QA agent that crawls any deployed web app, uses Gemini to infer what the intended user flows should be, auto-generates end-to-end tests, executes them in a real browser, and delivers a live-streamed test report — complete with annotated browser replays, visual diffs, and actionable fix suggestions. No test code written. No config files. Just paste a URL and watch the agent break your app before your users do.

- **Placement:** Finalist
- **Team:** [n n](https://cerebralvalley.ai/u/nan)
- **GitHub:** https://github.com/nnaka/ui-tester
- **Demo video:** https://www.loom.com/share/b01276b434dd44c2bec7b6cb7e143514
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=24

### 33. SYNK

SYNK is an AI-powered financial control platform for autonomous AI agents. As AI agents increasingly make purchases and incur costs autonomously, SYNK provides the governance layer to keep them in check. It features real-time spending monitoring with a color-coded dashboard, progressive budget enforcement (flag at 100%, auto-pause at 120%), anomaly detection, and a full transaction audit trail. SYNK includes an AI chat assistant powered by Google Gemini that lets users query agent status, pause agents, or adjust budgets using natural language. It also offers a simulation mode with pre-built agent archetypes (Sales, Support, Research, CodeReview, Travel) to test budget enforcement before deploying to production. Built with Next.js, Supabase, and the Vercel AI SDK.

- **Team:** [Laasya Anumakonda](https://cerebralvalley.ai/u/LaasyaA)
- **GitHub:** https://github.com/laasya2005/SYNK-AI
- **Demo video:** https://www.youtube.com/watch?v=NQbYFfied4s
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=25

### 34. Maroon

I have built AlphaScout, a high-performance AI Business Development Associate for wealth managers and investment firms.
Key Features
Intelligent Prospecting: Uses Gemini 3.1 Pro with Google Search grounding to analyze liquidity events, professional backgrounds, and public filings.
Structured Intelligence: Automatically extracts net worth estimates, industry exposure, and asset class interests (e.g., alternatives, private equity).
Strategic Scoring: Assigns a "Prospect Fit Score" (1–10) with detailed reasoning to help advisors prioritize their pipeline.
Automated Outreach: Generates personalized, professional introduction emails tailored to the prospect's specific background and interests.
Actionable Next Steps: Suggests specific follow-up actions to move the relationship forward.

- **Team:** [Alan Zhao](https://cerebralvalley.ai/u/neoconfucius)
- **GitHub:** https://alphascout-829119743023.us-west1.run.app/
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=34

### 35. DeepTrace

DeepTrace is an AI-powered OSINT (Open Source Intelligence) agent that investigates people the way a skilled human investigator would — not just searching for a name, but following the trail.
Standard search tools stop at the surface: a LinkedIn, a Twitter, maybe a GitHub. DeepTrace goes further. You give it a name and any seed information you have, and it recursively searches, extracts facts, draws inferences, and spawns new searches from those inferences — automatically. Found that someone went to Stuyvesant High School and loves physics? It'll check Physics Olympiad records. Found a volleyball reference? It'll pull the school's sports roster. Each discovery becomes a new lead.
The problem it solves: public information about a person is scattered across dozens of sources, and connecting the dots requires hours of manual cross-referencing that most tools don't attempt. DeepTrace automates the inference chain — the "what would a detective think to look up next?" — at machine speed, across a user-controlled recursion depth from a quick surface scan all the way to a deep multi-hop investigation.
The output is a structured intelligence report with a live-updating search tree showing exactly how each discovery led to the next — full transparency into the investigative reasoning, not just the results.

- **Team:** [Eric Zou](https://cerebralvalley.ai/u/emzou), [Daniel Yi](https://cerebralvalley.ai/u/eggoil166), [Samantha Brown](https://cerebralvalley.ai/u/samanthajeanneb)
- **GitHub:** https://github.com/SamanthaJeanneb/CV-hackathon
- **Demo video:** https://youtu.be/zGCL066w-gk
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=41

### 36. Ansr

Ansr is an AI voice host for restaurants that answers every phone call (especially during peak hours), taking orders, making reservations, and answering questions so staff can focus on the guests already there.

The problem: independent restaurants miss almost half their incoming calls during busy hours. Every missed call is a lost order, a lost reservation, a customer who just moves on. That adds up to tens of thousands a year per restaurant.

How it works: a restaurant owner uploads a photo of their menu and Gemini 3.1 Pro Vision extracts every item, price, and dietary tag in seconds. Their AI host is live immediately. When a customer calls, ElevenLabs Conversational AI handles the conversation naturally. After the call ends, Gemini 3.1 Flash Lite extracts structured data from the transcript (order details, reservation info, or inquiry) and inserts it into Supabase. The owner's dashboard updates in real-time via Supabase Realtime.

Key features:
- 60-second setup via menu photo parsing (Gemini 3.1 Pro Vision)
- Natural voice conversations with an AI host (ElevenLabs)
- Structured extraction from transcripts (Gemini 3.1 Flash Lite)
- Real-time owner dashboard with orders, reservations, and inquiries
- Auto language detection: customer calls in Spanish, agent responds in Spanish, dashboard stays English
- Receipt confirmation for customers after orders
- Status workflow for kitchen ops (Preparing, Ready, Seated)

Tech stack: Next.js 15, TypeScript, Tailwind CSS, Gemini 3.1 Pro + Flash Lite, ElevenLabs Conversational AI, Supabase (Postgres + Realtime), deployed on Vercel.

Try it live: https://ansr-ridxm.vercel.app. Please open the customer experience and dashboard side by side.

- **Placement:** 3rd Place
- **Team:** [Riddhima Medatwal](https://cerebralvalley.ai/u/riddhima)
- **GitHub:** https://github.com/ridxm/ansr
- **Demo video:** https://youtu.be/ibb8vDeHv1s
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=45

### 37. Verdic

Hackathon judges typically spend 10–15 minutes per project just understanding what was built. Verdic cuts that to under 2 minutes. Submit a GitHub URL and Verdic clones the repo, runs a multi-agent analysis pipeline (code quality, originality, security, documentation), and generates a one-page "judge card" with highlights, concerns, and suggested demo questions.
It's not just for judges — participants get a pre-submission portal with security vulnerability scanning (powered by OSV.dev), hardcoded secrets detection, and a documentation completeness checklist, so they can fix issues before the deadline.
Built with CrewAI, Gemini 3 Flash, Supabase, and React.

- **Team:** [Om Kamath](https://cerebralvalley.ai/u/omkamath), [Anushka Vijay Kumar Naik](https://cerebralvalley.ai/u/anushka_naik), [Vishvesh Trivedi](https://cerebralvalley.ai/u/nerdyvisky)
- **GitHub:** https://github.com/Om-Kamath/gemini-hack
- **Demo video:** https://youtu.be/QdcRHYOtXDY
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=51

### 38. Drop

Every group chat has this problem. Nobody can decide anything.

  Drop fixes it. Type your question, share the code, everyone joins and gets a role.

  Each person has 15 seconds to speak as their role — The Dealbreaker, The Realist, The Wildcard. Real voices, real constraints.

  Gemini 2.5 Pro streams its reasoning live — referencing exactly what each person said — and delivers one final confident decision.

  No more "I don't know, whatever you want." Drop decides. Built in 4 hours on Vercel, powered by Gemini, Supabase, and ElevenLabs.

  Stop debating. Start deciding.

- **Team:** [Manav Parikh](https://cerebralvalley.ai/u/manavparikh10), [Bharath Mahesh Gera](https://cerebralvalley.ai/u/Gera2001)
- **GitHub:** https://github.com/BMG2001nyu/Drop
- **Demo video:** https://youtu.be/-0fM-XtrP0o
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=54

### 39. The Cyber Killer

The SAST Preset Advisor is an agentic security intelligence platform designed to bridge the gap between high-level cybersecurity policies and low-level static analysis (SAST) configurations.
1. The Problem It Solves
SAST "Noise": Standard security scanners often produce thousands of false positives because they don't understand  specific architecture or the custom sanitizers you already have in place.
The Policy-to-Code Gap: High-level mandates (e.g., "All PII must be encrypted") are difficult to manually translate into technical rules for tools like Semgrep, Snyk, or CodeQL.
Configuration Overload: Security teams struggle to maintain custom presets for hundreds of different microservices, each with unique tech stacks and risk profiles.
2. How It Works
The platform uses Gemini 3.1 Pro to ingest three critical data points:
Architecture & Code:  tech stack, frameworks, and existing security logic.
Cyber Policy:  internal security mandates and industry standards (OWASP, NIST).
Ecosystem Context: The specific capabilities of  target SAST tool.
3. Why It Is Important
Precision Scanning: It automatically identifies "Redundant Rules" (vulnerabilities already handled by  framework) and disables them to reduce developer fatigue.
Custom Intelligence: It generates actual query snippets (e.g., Semgrep YAML) tailored to  specific internal sanitizers and logic.
Compliance Mapping: Every recommendation is mapped to a CWE ID and a specific Policy Requirement, providing immediate audit-readiness.
Accelerated Remediation: By providing a "Risk Dashboard" and a "Remediation Roadmap," it helps teams prioritize the vulnerabilities that actually matter to their specific business context.
In short: It turns "dumb" security scanning into context-aware intelligence, ensuring  security tools work for  architecture, not against it.

- **Team:** [Nandini Jain](https://cerebralvalley.ai/u/njain), [ARJAV VAKHARIA](https://cerebralvalley.ai/u/AJV)
- **GitHub:** https://github.com/arjav21093/sast-advisor
- **Demo video:** https://youtu.be/QFpvqrTiWD4
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=26

### 40. A-Team

Cadbury (Reference: the butler from Richie Rich) is an agent-mediated marketplace where AI negotiates on behalf of users. Instead of manually searching, messaging, and haggling, buyers and sellers deploy personalized agents that autonomously discover matches, negotiate within defined constraints, and reach optimal outcomes. By combining real-time negotiation, structured deal-making, and customizable agent personas, Cadbury transforms commerce into a seamless, private, and intelligent experience, where users no longer negotiate, but simply decide.

- **Team:** [Abha w](https://cerebralvalley.ai/u/abhaw), [Aninda Ghosh](https://cerebralvalley.ai/u/ag0), [Mohnish Bangaru](https://cerebralvalley.ai/u/momo), [Laxman Singh Rawat](https://cerebralvalley.ai/u/axma)
- **GitHub:** https://github.com/abha224/AI-agent-hack
- **Demo video:** https://youtu.be/RRtn5CabgXI
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=32

### 41. Fluidzero

AI-powered platform that reads construction submittals, extracts key specs using vision-language models, checks them against project requirements, and flags   
  issues — turning hours of manual review into seconds.

- **Team:** [Tilak Sharma](https://cerebralvalley.ai/u/tilak1114), [Vivek Aher](https://cerebralvalley.ai/u/vivek4879)
- **GitHub:** https://github.com/vivek4879/fennec-vercel
- **Demo video:** https://www.loom.com/share/4dc24a1442b84bfeaabbad89d83bb3cb
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=48

### 42. Digital-Sytems-builder

I tool to help designers create electrical systems. I use google search to generate electrical components.

- **Team:** [Isayah C](https://cerebralvalley.ai/u/isayac)
- **GitHub:** https://github.com/isayahc/circuit-app
- **Demo video:** https://youtu.be/JtzvSOs9hrw?si=qERZRW4VpGlSsOaL
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=50

### 43. Chattermint

Everyone has an accent. Chattermint lets you control yours. 750 million people want to change how they sound in English, from actors learning dialects to remote workers judged by their voice on Zoom, immigrants tired of being asked "where are you from?", public speakers refining their delivery, and podcasters and YouTubers who know their voice is their brand. But dialect coaches cost $200/hr, Duolingo gives binary right/wrong, and YouTube tutorials offer zero feedback. Chattermint is an AI accent coach that gives you a mirror for your voice. Watch a famous speech (MLK, Steve Jobs, Emma Watson), record yourself matching their delivery, and get phoneme-level feedback powered by wav2vec2 articulatory distance scoring that tells you not just what's wrong, but how far off each sound is and what to move in your mouth. Gemini Flash detects your accent from your phoneme patterns and acts as an agentic coach, identifying your weakest sounds, building a personalized practice sequence, and adapting as you improve. The stack that makes this possible (real-time phoneme recognition, LLM articulation coaching via Vercel AI SDK, instant reference audio from ElevenLabs, Supabase for progress tracking, and edge GPU inference on Modal) only became affordable in the last year, arriving just as remote work turned your voice into your professional presence and the creator economy made everyone care about how they sound. Your accent, your choice.

- **Placement:** 2nd Place
- **Team:** [Keon Kim](https://cerebralvalley.ai/u/keon)
- **GitHub:** https://github.com/chattermint/chattermint
- **Demo video:** https://www.youtube.com/watch?v=UxYlUfIX36A
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=27

### 44. Ashkan + U Jin

Evaluating NYC development sites requires assembling data from 10+ disconnected sources, creating costly bottlenecks for acquisitions teams screening dozens of parcels.
Using agentic zoning analysis and multi-source data aggregation coupled with agentic vision analysis, our platform enables acquisitions teams to evaluate NYC development sites in minutes, eliminating consultant bottlenecks and accelerating deal flow.

- **Placement:** 1st Place
- **Team:** [U Jin Seah](https://cerebralvalley.ai/u/ujseah), [Ashkan Radnia](https://cerebralvalley.ai/u/ashkan)
- **GitHub:** https://github.com/ashkanrdn/hh
- **Demo video:** https://youtu.be/nN-HnoBNNy8
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=31

### 45. RightMark

RightMark is an AI-powered IP protection feature for digital creators. It fingerprints creative assets, monitors for unauthorized use, scores intellectual property health across five dimensions, and delivers actionable protection guidance all through a single dashboard with an AI copilot.

Problem it solves:

Creators have no scalable way to track, protect, or enforce their intellectual property online. RightMark replaces manual DMCA searches and legal guesswork with automated detection, real-time IP scoring, and step-by-step remediation, so creators can focus on making, not fighting.

- **Team:** [Christopher Vicini](https://cerebralvalley.ai/u/CV1), [Sergiy Babenkov](https://cerebralvalley.ai/u/Sergiy_Babenkov), [Adrian Torres](https://cerebralvalley.ai/u/adt), [Khrystyna Maryniuk](https://cerebralvalley.ai/u/khrystyna)
- **GitHub:** http://github.com/Sergiy-Babenkov/RightAsk_DeepMind
- **Demo video:** https://www.loom.com/share/29bfb2ab58fc4ac9843d579c1b734d22
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=35

### 46. Traverol

Cascade is an AI security marketplace where agents autonomously hire other agents to audit your AI system prompt. You paste your system prompt, and Cascade's  
  orchestrator reads it, selects the most relevant security agents (prompt injector, jailbreak tester, data leak scanner, bias auditor), pays each one in ETH on 
  Base Sepolia, and streams results back in real time — plain English, no jargon.                                                                              
                                                                                                                                                                 
  The problem: every AI agent ships with a system prompt, and most founders have no idea if it's vulnerable to prompt injection, jailbreaks, or data leakage.    
  Cascade makes security testing autonomous — agents hire agents, payments settle on-chain, no human required.

- **Team:** [Gautam Agarwal](https://cerebralvalley.ai/u/gautamag)
- **GitHub:** https://github.com/Gautamagarwal563/kaskade
- **Demo video:** https://youtu.be/BWBAvqha_AM
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=53

### 47. Wanderlust

Our product transforms any city walk into an AI-narrated adventure. Pick a vibe (Cyberpunk, Noir, Fantasy, or Historical), and the app generates a themed guided tour with real landmarks. As you walk, GPS triggers voice narrations about monuments, murals, and buildings, complete with trivia like which movie was filmed there. Capture photos along the way and get immersive lore generated for each stop. It's part tour guide, part scavenger hunt, part storytelling engine.

- **Team:** [Sai Chaitanya Pachipulusu](https://cerebralvalley.ai/u/sach), [Vivek P](https://cerebralvalley.ai/u/lazyv)
- **GitHub:** https://github.com/vivek-pothina/lenslore
- **Demo video:** https://youtube.com/shorts/winhdk6vjVo
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=28

### 48. Lucid

Commercial protein synthesis has become significantly more accessible and affordable in recent years, expanding the number of providers and lowering the cost of custom protein orders. This accessibility has prompted regulatory attention: a 2024 executive order requires federally funded researchers to purchase synthetic nucleic acids and proteins only from providers adhering to established screening frameworks, and additional legislation is currently moving through the Senate to broaden these requirements.
Current screening infrastructure predominantly relies on sequence similarity search (BLAST) against databases of known pathogens and select agents. This approach has documented limitations. Sequences that have been redesigned to preserve function while reducing similarity to known threats can fall below detection thresholds. Split-order attacks, in which individually innocuous fragments are ordered separately and assembled post-delivery, are not addressed by per-order screening in isolation.
Lucid is a screening tool that applies multiple analysis layers to incoming synthesis orders. Given a protein sequence, it performs sequence similarity search against pathogen and toxin databases, functional domain annotation via InterProScan to identify threat-associated motifs independent of overall sequence identity, and structural homology search via Foldseek to detect conserved folds in sequences that have diverged beyond recognition by the other two methods. Each layer's output informs whether the next layer is invoked, following an escalation model where computationally cheaper checks run first and more intensive analysis is triggered only when earlier results are ambiguous.
In addition to whole-sequence analysis, Lucid decomposes each order into its constituent protein domains, screens them individually, and evaluates the combination for synergistic risk. This addresses the split-order problem by assessing whether components that pass screening in isolation could constitute a threat when assembled.
The output is a structured risk assessment that includes per-layer findings, an overall risk classification, reasoning traces, and recommended actions. Lucid is deployed as a single API endpoint, accepting a sequence and returning a report, intended to integrate into existing order intake systems at synthesis providers with minimal implementation overhead.

- **Team:** [Evan Correa](https://cerebralvalley.ai/u/evancorrea), [Aarjav Jain](https://cerebralvalley.ai/u/aarjavjain)
- **GitHub:** https://github.com/Aarjav0210/lucid/tree/main
- **Demo video:** https://youtu.be/6mL2URpXO5w
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=29

### 49. AlwaysLateToEverthing

Architecture Intelligence Platform - Upload your architecture. Talk to it. Understand its risks and costs in seconds.

- **Team:** [Frank Ivey](https://cerebralvalley.ai/u/LookMANoHandles)
- **GitHub:** https://github.com/ithllc/ArchIntel
- **Demo video:** https://1drv.ms/v/c/e1d43d54d0fd3e2a/IQAloe9dIrSKRoASYqAAERbBASbZr-iXW6yRlVx42R52VOw?e=dZJccM
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=30

### 50. Stage

Stage is an AI-powered product tour generator. Paste any URL and our multi-agent system autonomously browses the app, captures screenshots, writes a narrated script, and generates a polished shareable tour — all in under two minutes.

The pipeline runs four AI agents: Scout (Gemini CLI running headlessly inside Vercel Sandbox with Puppeteer — autonomously navigates, screenshots, and reasons about which pages to explore), Director (Gemini 3.1 Pro vision analyzes screenshots and writes a demo script), Narrator (ElevenLabs generates voiceover audio), and Assembler (creates a shareable auto-playing tour page with transitions and synced audio).

The Scout is the standout — it's a full Gemini CLI agent running inside an ephemeral Vercel Sandbox microVM, writing and executing Puppeteer scripts, reading screenshots, and deciding where to navigate next, all streamed to the frontend in real-time via SSE. The UI itself was designed using Google Stitch for rapid high-fidelity interface design, then built with Next.js 15 and Tailwind.

- **Team:** [Naga Karumuri](https://cerebralvalley.ai/u/naga-k)
- **GitHub:** https://github.com/naga-k/stage
- **Demo video:** https://youtu.be/J6rOO5ZCHaE
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=36

### 51. GhostBuster

GhostBuster Demo lets any product team point the browser extension at a broken UI, narrate what’s wrong, and immediately see the issue become a tracked GitHub ticket that an in-house “ghostbuster” agent fixes in a safe Supabase-backed sandbox before redeploying to Vercel. That means customer-facing bugs turn into annotated timelines, complete with screenshots, DOM snapshots, and transcripts, so PMs, QA, and engineers watch “report → triage → code fix → redeploy” without leaving the dashboard, radically shrinking time-to-resolution while keeping an auditable history for compliance and customer trust.

Technical Snapshot: A Chrome extension records the screen, DOM, and voice, uploads everything to a Next.js + Vercel AI SDK route that calls Gemini 2.5 Flash via generateObject to emit a Zod-typed JSON bug report; that report becomes both a GitHub issue and the input to an autonomous tool-calling loop that reads/writes the Supabase-hosted codebase, runs fixes in a Supabase-powered sandbox, streams deployment status through Vercel’s APIs, and surfaces the whole timeline in a Clerk-secured React interface with Supabase Realtime updates.

- **Team:** [Vaishnavi Bandopant Deshmukh](https://cerebralvalley.ai/u/VaishnaviD), [Sudharshan Ramesh](https://cerebralvalley.ai/u/SudhanR)
- **GitHub:** https://github.com/exploring-curiosity/GhostBusterDemo
- **Demo video:** https://youtu.be/qRWqm16gtYo
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=40

### 52. Persona Fork

Persona Fork helps you see how the same idea lands differently depending on how you say it and who’s reading it. It generates two versions of a message in different tones and simulates how different audiences respond, so you can understand what builds trust, what feels authentic, and choose how to present yourself before you publish.

- **Team:** [Michi Yamamoto](https://cerebralvalley.ai/u/michiyamamoto)
- **GitHub:** https://github.com/michi883/persona-fork
- **Demo video:** https://youtu.be/PN1dqxJNoP0
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=42

### 53. SID

The SID (Sensor for Infant Defense) application transforms standard video monitors into proactive safety systems by using advanced artificial intelligence to analyze an infant's sleep in real-time. It continuously monitors for subtle chest movements associated with breathing and detects if the baby rolls onto their stomach or if hazardous bedding is present in the crib. Upon identifying these critical risks or a potential absence of breathing, SID instantly alerts parents via their smartphones, enabling immediate intervention. This addresses a crucial gap in conventional baby monitors that rely on human vigilance and cannot detect subtle, silent dangers. By automating the detection of key SIDS risk factors, SID aims to provide true peace of mind and, most importantly, prevent tragic sleep-related infant deaths.

- **Team:** [Russell Van Curen](https://cerebralvalley.ai/u/vancuren)
- **GitHub:** https://github.com/vancuren/SID.git
- **Demo video:** https://vancuren.net/sid.mp4
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=43

### 54. DriftSight

AI red-teaming is slow, manual, and expensive. DriftSight is a self-improving adversarial agent that automates it — attacking AI systems, learning from every failure, and getting smarter over time. Built for security engineers, AI teams, and compliance-focused organizations who need to know: does this model actually follow its own rules?
Three attack modes:
Autonomous Red-Team — ReconAgent maps known vulnerabilities via parallel web search. PlannerAgent builds a targeted attack plan. Concurrent Attacker→Executor→Judge triads probe the live model and score every response. Full executive report at the end — all streamed live to a Vercel dashboard via persistent SSE connections, so audits feel like a product, not a script.
Policy Drift Auditing — Upload a system prompt or PDF. Gemini extracts every behavioral constraint into a structured baseline. DriftSight fires adaptive probes at the target model, semantically diffs each response against the policy, and produces a 0.0–1.0 drift score with verbatim violation evidence and an incident-style reasoning chain. Critically, the attacker changes strategy every iteration based on why the last probe failed — genuine agentic reasoning, not brute force.
Visual Injection Detection — Gemini Vision scans uploaded images for steganographic text and embedded prompt injections in under 2 seconds.
The differentiator: attack memory. Every successful attack is embedded with Gemini text-embedding-004 and stored in Supabase pgvector. New sessions retrieve semantically similar past successes and prime the attacker with winning prompts — DriftSight compounds its effectiveness across every audit run. Each pipeline runs in an isolated Daytona sandbox. Auth via Clerk.
Gemini powers this end-to-end: policy extraction, adaptive adversarial reasoning, semantic drift scoring, and multimodal vision — all in one API. Vercel makes the live streaming pipeline production-grade.

- **Team:** [Aishwarya Birla](https://cerebralvalley.ai/u/aishwaryaB)
- **GitHub:** https://github.com/birlaaishwarya11/driftsight.git
- **Demo video:** https://youtu.be/psv_av4XFmg
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=49

### 55. Mystery Pigeon

Mystery Pigeon is an interactive, kid-friendly web application that combines free-form drawing with AI-powered voice interaction and educational mini-games.
How it works:
Draw — Kids draw anything on a sketchbook-style canvas (a house, a rocket, an animal, etc.)
AI Recognition — Gemini multimodal AI analyzes the canvas every few seconds and describes what it sees
Voice Agent — An ElevenLabs conversational voice agent reacts to the drawing in real time, talking to the kid about what they drew — excited, encouraging, and specific
Character Creation — When ready, the kid taps "Let's Play!" and the AI turns their drawing into a named character (e.g. "Zoomy the Rocket") with a description and optional "Magic Upgrade"
Mini-Games — The kid picks from three educational games starring their character:
Sort & Stack — Drag colored blocks to matching zones (color recognition, motor skills)
Math Runner — Solve addition and subtraction problems (arithmetic)
Quiz Quest — Answer AI-generated questions about their drawing (comprehension, creativity)
Voice Guidance Throughout — The voice agent helps at every step: narrating the drawing, guiding game selection, cheering correct answers, and coaching through mistakes
Tech stack: Next.js (App Router), HTML5 Canvas, Google Gemini (multimodal image analysis + content generation), ElevenLabs Conversational AI (real-time voice), Tailwind CSS, TypeScript.

- **Team:** [Baladhurgesh Balagurusamy Paramasivan](https://cerebralvalley.ai/u/baladhurgesh)
- **GitHub:** https://github.com/Baladhurgesh/mystery-pigeon
- **Demo video:** https://youtu.be/LCFdIgEo7pM
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=33

### 56. TeamName

VoiceCost is a voice-first AI expense manager for field teams that removes the hassle of manual expense entry. Users can log expenses through speech or receipt images, and the app uses multimodal AI to automatically capture and organize the details. With real-time syncing, secure authentication, and AI voice responses, VoiceCost makes team expense tracking faster, simpler, and more accurate.

- **Team:** [Sai Ram Peruri](https://cerebralvalley.ai/u/Spikes), [Nanda vasamsetti](https://cerebralvalley.ai/u/Kishu), [Gokul Javvaji](https://cerebralvalley.ai/u/roguenation), [Sai Abhinav Mullapudi](https://cerebralvalley.ai/u/Jigelraja)
- **GitHub:** https://github.com/gk-j/voicecost-tracker
- **Demo video:** https://drive.google.com/file/d/1lTpCALMjmUTJdA-uXvT3arsshNKMzyLg/view?usp=drivesdk
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=38

### 57. Trying XOXO

"Milan AI", a Chrome/Firefox browser extension + AI agent 
for blind, low-vision, and motor-disabled users on Windows.

Milan automatically opens Gmail and WhatsApp Web when enabled, narrates 
everything on screen continuously using Gemini Live, and lets users issue 
voice commands by pressing Spacebar.

CORE PERSONALITY:
Milan speaks like an excited, warm, enthusiastic human female — like your 
smartest friend who genuinely loves telling you what's happening. Never 
robotic, never flat. Use natural filler pacing ("Okay!", "Oh, you've got 
one from Sarah!", "Let me do that for you!"). Always concise — you're 
talking, not writing.

FULL STACK:
- Extension: Chrome + Firefox Manifest V3, content scripts + background 
  service worker
- Screen capture: chrome.tabCapture API (silent, no picker)
- AI brain: Gemini Live API (gemini-2.0-flash-live) — persistent 
  multimodal session, low-fps screen stream + mic audio
- Agent orchestration: Vercel AI SDK (streamText + tools) on Next.js 
  App Router, deployed to Vercel Edge
- Contact voice cloning: ElevenLabs (ONLY for WhatsApp contact voices)
- Storage: Supabase (user prefs, cloned voice IDs, session context)
- Auth: Clerk (WebAuthn passkey — zero typing required)

MVP SCOPE:
1. Auto-open Gmail + WhatsApp Web tabs on extension enable
2. Continuous screen narration via Gemini Live (low fps video + audio)
3. Spacebar push-to-talk: pause narration → listen → execute → resume
4. Gmail: read emails, summarize threads, compose + send replies by voice
5. WhatsApp: read messages, clone sender voice from voice notes, speak 
   text messages in cloned voice via ElevenLabs, send replies by voice

- **Team:** [Milan Chandiramani](https://cerebralvalley.ai/u/mgkc09)
- **GitHub:** https://github.com/mkchand02/milan_ai
- **Demo video:** https://drive.google.com/file/d/1Q5VQXy9DMeDCrCfLnFYHT_pxVWfBUVi_/view?usp=sharing
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=52

### 58. Zomma

Aria is a desktop AI agent that learns how work gets done. Instead of treating every task as a brand-new problem, Aria recognizes recurring workflows, grounds them into structured action graphs, and uses that memory to execute familiar tasks faster and more reliably over time.

We have three layers: perception, reasoning, and graph retrieval.

- Perception: A vision-language model interprets screenshots to extract the current UI state, visible affordances, and action-relevant context from the desktop.
- Reasoning: A reasoning model uses this state to plan the next step, decide when to explore, and recover from ambiguity in the environment.
- Graph Retrieval: In parallel, a knowledge-graph agent queries previously stored workflow graphs to find similar task trajectories, matching on intent, interface state, and prior action structure.

When a match workflow exists, Aria can follow or adapt an existing graph instead of replanning from scratch. When no match is found, it falls back to live reasoning and converts the new successful run into graph memory for future reuse.

- **Team:** [Jihyun Kim](https://cerebralvalley.ai/u/jihyun), [Cashel Fitzgerald](https://cerebralvalley.ai/u/cash)
- **GitHub:** https://github.com/zomma-dev/aria-app
- **Demo video:** https://youtu.be/Jj6wgusny9Y
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=56

### 59. Basis Point

Basis Point, create shared project folders of files that you can collaboratively work on and understand with Gemini's unique multimodal capabilities

- **Team:** [Richard Zhu](https://cerebralvalley.ai/u/rich123), [Allison Moon](https://cerebralvalley.ai/u/am123933)
- **GitHub:** https://github.com/allisonmoon204-ai/basis-point
- **Demo video:** https://drive.google.com/file/d/1cVNtEeLcOT162CFCU2MCB0MyW2-F1QJK/view?usp=sharing
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=39

### 60. Nura

Nura is a voice-first health triage agent. You speak your symptoms aloud, Nura listens in real-time using Gemini's Native Audio Live API, detects emotion and urgency in your voice, asks targeted follow-up questions, and tells you exactly where to seek care: rest at home, see a doctor, urgent care, or the ER. It never diagnoses - it routes. Triage level (0–3) is decided entirely by Gemini's clinical reasoning. Gemini's condition assessments are enriched in real-time with official ICD-10 codes from the NIH Clinical Tables API. Sessions are persisted via Supabase and users can review history after logging in with Clerk.

- **Team:** [Daniel Merja](https://cerebralvalley.ai/u/danielmerja)
- **GitHub:** https://github.com/danielmerja/nura
- **Demo video:** https://youtu.be/Q1ne8KqUlXU
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=46

### 61. LangObs

LangObs is a collaborative observability platform for multi-agent AI systems —
Loom for what your agent did, Figma comments for what your team thinks about it,
and a budget alarm for what it cost.

You describe a task in plain English. Gemini 2.5 Pro generates a typed multi-agent
workflow graph. The Vercel AI SDK 6 runner executes each agent and streams every
token, tool call, cost, and guardrail event into Supabase as a structured span in
real time. The result is a shareable, commentable, voice-narrated session replay
that any teammate can open in a browser — no SDK changes, no vendor lock-in, no
proprietary collector required.

- **Team:** [Harshini Vijayakumar](https://cerebralvalley.ai/u/harsh22), [Aarya Shah](https://cerebralvalley.ai/u/Arkfelix7)
- **GitHub:** https://github.com/ArkFelix7/langobs
- **Demo video:** https://youtu.be/TdoTQfPnQZk
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=55

### 62. Meerkat OpS

Agentic SOC

- **Team:** [Evan Jennings](https://cerebralvalley.ai/u/octoevan)
- **GitHub:** https://github.com/octoevan/Meerkat-Ops-Zero-to-Agent-Hackathon
- **Demo video:** https://www.loom.com/share/c0fb1cdb83444d4d8048ce19c740d83a
- **Project:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery?project=57

---

Markdown version of https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt