GPT-5 Startup Hackathon NYC

Bellhop is an AI Chief of Staff for boutique hoteliers—right in their pocket. Running a boutique hotel is complicated. Hoteliers are on their feet all day, focused on guests and operations. They don’t have the time—or the tools—to manage pricing, marketing, and daily tasks from behind a desk. Enter Bellhop: an AI co-pilot that automates revenue management and marketing, surfaces key insights, and empowers hoteliers to interact through a simple chat or voice interface. By increasing revenues and saving valuable time, Bellhop lets hoteliers focus on what matters most: their guests.

Fathom is a voice first, AI powered codebase assistant that eliminates the code comprehension barrier for developers. Think of it as the first Siri for codebases, enabling developers to speak naturally and get deep, actionable insights about their systems. It helps teams understand complex repositories faster using graph traversal, semantic chunking, and precomputed indexing. Developers can ask questions like “How does authentication work?” and get complete, real time answers. Fathom supports onboarding, debugging, and refactoring by treating the codebase as a living knowledge graph. It delivers context aware, conversational insights, remembers prior queries, and supports multilingual interaction, making it easier for non native English speakers to explore and understand codebases. It is not just code intelligence, it is code and infrastructure intelligence in one place to explore, query, and act on everything that makes the system run.

It's a way to maximize ROI on your credit cards...with a finance bro expert helping you out (aka me).

Uses gpt-5 and gpt-5-codex to find and refine unstructured renal dosing guidelines into python encoded clinical rules that then are served at the point of prescribing via HL7 compliant cds-services. There are between 250,000 and 300,000 deaths per year in the US due to Adverse Drug Events. This is a small step towards fixing that.

Asaply is a lightweight checkout engine that uses GPT-5 to translate natural language into structured orders. Customers describe what they want in plain English, and Asaply interprets the request, matches it to a merchant’s menu, verifies uncertainties, and generates a ready-to-process checkout plan. This makes ordering as easy as talking, while saving merchants the need to build custom chatbots or rigid menus.

Tesseract is Devops first coding space that lets you build and modify and deploy webapps with a canvas like design

Minam is the platform that democratizes API creation for knowledge creators. Transform data, insights, and expertise into tiered APIs that pay you every time someone accesses your knowledge. Control who gets access and how much - set up tiers, fractionalize your data, and get paid for what you need with a pay-as-you-need model. Just like OnlyFans and YouTube revolutionized content creation, Minam will revolutionize API creation for knowledge creators.

The Anote Model Leaderboard is a public API-driven benchmarking platform for evaluating large language models (LLMs) across diverse task types. It solves the problem of fragmented, slow, and manual model evaluation by providing a scalable, transparent, and community-driven system where new datasets and model submissions can be added programmatically. The platform supports multiple task categories including: Text Classification Named Entity Recognition (NER) Document-level Q&A (chatbot) Line-level Q&A (prompting) Researchers and enterprises can benchmark zero-shot, fine-tuned, and supervised models. This ensures fair comparison of models, and encourages reproducibility through standardized evaluation metrics.

- We built the ContextCity Codex concierge: an Expo mobile experience where Dartagnan guides travelers through city highlights. - GPT-backed workers curate social signals into reels, Remotion renders localized MP4s, and the app surfaces rationale, itinerary, and concierge cues. - A “Route to {POI}” card and localized overlay tie reels directly to multi-step actions (plan, book, guide, celebrate). - The demo showcases Felix in SoHo across French, English, and Spanish, including synchronized narration, subtitles, and concierge voice.

Help the English learner to practice the oral and listening English; Ace the TOEFL Exam.

The GPT Negotiator Agent is an AI-powered financial & legal assistant for the music industry, designed to bridge the gap between creators and the institutions that manage their royalties, licensing, and financing.

For b2b AI companies to measure their visibility in LLM. Generate brand-visibility briefs by masking provider names and comparing how multiple LLMs (GPT-5 primary, GPT-4o comparison) respond. The pipeline ingests a transcript, extracts selling points, asks domain-expert questions, and highlights where models still infer the masked provider.

Glimpse into a world without manual data entry. We fit any data you upload into any template of your choice.

Memory Playground turns your everyday chats into things you actually want to look at: a front‑page newspaper of your life, a lovingly savage “Roast My Week,” tap‑to‑play trivia, a 3‑panel comic illustrated from your prompts, and advice from “Future You.” It’s a personal highlight reel that feels like entertainment, not homework. You open the app and you’re the main character: headlines, jokes, panels, and mini‑games built from your own messages and voice notes. Why it matters: reflection is hard; scrolling is easy. Memory Playground sneaks reflection into micro‑delights—funny panels, quick games, and shareable moments—so people come back for the smiles and stay for the insight.

Team Humor Genome is building Meme Lab, an app that analyzes and remixes memes to explore different humor styles. Meme Lab uses AI to label memes, generate creative edits, and collect user feedback to build a personal "humor fingerprint." The project is an early prototype of how AI can learn and adapt to what makes people laugh.

Luxury hospitality is booming, with Ritz, Marriott, and others betting big on branded residences and timeshares. Yet booking still feels like 2005: static forms, no context, and concierge stuck in email. Spotlight Concierge changes that. Guests select dates, building, floor, and apartment, then an AI concierge personalizes the journey: suggesting up relevant services, capturing custom requests, and rewarding engagement with surprises like complimentary champagne. It makes digital booking feel like true hospitality. With branded residences projected to reach $70B+ by 2031 at 11-16% annual growth, and services already driving 10–15% of hotel revenue, Spotlight helps operators capture that upside by turning lost upsell moments into real value while giving guests the VIP experience they expect.

MergeFund is a GitHub-native bounty platform that lets open-source repositories attach real money to open issues. When a contributor submits a pull request (code) that resolves an issue, the payout is automatically released upon merge no delays, no manual steps. Maintainers, companies and Open-source repositories can accelerate bug fixes or feature development without changing their workflow, while developers get fairly compensated for their contributions. By aligning incentives and removing friction, MergeFund keeps open-source projects active, healthy, and sustainable.

Some smaller, specialized models can outperform larger models on specific tasks. What if we have enough coverage from these specialized models to outperform larger models on general tasks? This becomes a routing issue. I present autorouter, a publicly served API, compatible with the OpenAI Python SDK that you can use right now. Achieves state of the art.

What if onboarding flows to products were friendly playful and optimized themselves over time? Vibe Onboarding is an agent that talks to your users to onboard them so they don't have to click through a boring UI After the user authenticates, it runs an enrichment / agentic search to pull information about them that it uses to automatically fill in fields like Name, Workspace and personalize the conversation and suggest other fields like username or inviting collaborators. WORKING DEMO: https://screen.studio/share/h5koRc20?state=uploading CODE WALKTHROUGH: https://screen.studio/share/QEEo5aQN?state=uploading

Rescue is an iOS app that lets you schedule a convincing incoming call exactly when you want it. You pick who’s “calling” and why, and at the chosen time your iPhone rings with a short, natural, in-character conversation that stays on theme. It’s perfect for breaking the ice, escaping awkward moments, playful pranks, or getting a quick pep talk—fun, fast, and believable

Project by Syed Shah (Syedwshah.nyc@gmail.com) & Oleg Savchuk (Oleg.savchuk12@gmail.com). Venue Agents handles the backend logic that works alongside a voice-agent concierge: the agent operates on a workflow using realtime data and persistence, allowing users to making custom booking reservations with ease. The assistant interacts with a backend: it knows which rooms are free, locks one in, issues a door code, and saves a record the operator can verify, handles payment processing, and updates a database.

Quick Overview: PatentPilot is a multimodal, multi-agent co-pilot that turns raw invention disclosures (text + optional sketches) into attorney-ready patent draft packages in minutes rather than current slow process (~ 8 months to years) —and then follows the application through the full patent lifecycle and afterwards maitenance. We orchestrate specialized agents (Analyst → Skeptic → Judge → Aggregator → Artifacts Builder) with strict JSON contracts, per-step timeouts, and progress UI. Outputs include Abstract, Claims, Specification, LaTeX/Markdown, a Graphviz system diagram, a compiled PDF, and a ZIP bundle for counsel. It also human-in-the-loop, SaaS-like multimodal platform powered by GPT-5 and Codex that reimagines the $40B+ patent industry by making intellectual property protection faster, cheaper, and more accessible — while keeping legal oversight intact for full patent life cycle.

HealthInvest AI is a cutting-edge healthcare investment and consulting platform that leverages GPT-5's advanced reasoning capabilities

An end-to-end AI workflow where GPT-5 infers KPIs, Codex generates SQL/Python pipelines, and the Responses API orchestrates reporting-automating product analytics from PRD to polished report.

Ether AI is a voice agent and assistant that helps cloud engineers deploy AWS resources. It can help engineers study cloud infrastructure with a chat mode, and it automates different day to day on deployment, like JIRA ticket management and updating your team via Slack.

Esther, the todo list that does itself

The journey from a promising idea to an actual working prototype is long, involves weeks of research and slow development. Aspiring founders and VCs end up losing a lot of time conducting market research, building landing pages and spending money on ads to just to get their first signal of interest. This immense initial friction means many potentially great ideas are never pursued. VibeCheck is autonomous, multi-agent AI framework that transforms this weeks-long validation process into minutes. The user provides their raw idea in plain English. Our system then orchestrates a series of specialized AI agents to: 1) Refine the Idea 2) Performs a comprehensive market research 3) Generates and hosts a live landing page 4) Dynamically create 5) Creates dynamic AI personas

CampaignPulse: ViralIQ helps creators and marketers predict if their Reddit posts will go viral before publishing. It analyzes subreddit insights, generates campaign plans, and gives a viral score based on timeliness, novelty, clarity, and emotion. This saves time, avoids low-engagement posts, and boosts impact. https://github.com/shankasf/viral_checker

Continuous ocean monitoring to fuse live sensor data from different sources and data types to action live incident reporting.

Using GPT-5 and Codex, we built our Operations Cloud to help our enterprise partners make sense of their data in a practical, actionable way. Every day, as navNote mobile app is used on site—from the truck to the shelf, and from the shelf to the consumer—we capture valuable insights. Our platform doesn’t just store this data; it analyzes it, helping partners anticipate market trends, predict sales, and make smarter, faster decisions.

https://mini.gullie.io/ mini dynamic relocation planner! it will plan tasks, search the web to give context on your move, trigger housing searches and pull listings from zillow and help you structure your relocation, all controlled by a voice agent. It consolidates the entire moving process into one place.

ConvoFlow transforms networking conversations into high-quality, personalized follow-up emails using AI. The problem: generic follow-ups have 15% response rates because they're forgettable and impersonal. ConvoFlow uses GPT-5 to analyze conversation context, extract relationship signals, and generate professional emails that reference specific discussion points and shared connections. This creates emails that feel like natural conversation continuations, boosting response rates to 60%+ while saving professionals hours of manual email crafting.

A comprehensive hackathon-ready personal healthcare discovery platform that combines provider/resource search with AI-powered recommendations (similar to Pulse) multilingual and voice/SMS included for digital literacy. Built with Node.js, Express, Supabase, and OpenAI integration.

GlowingStar.AI reimagines how teams connect with customers by bringing emotional intelligence into every interaction. Today, digital conversations often miss subtle cues—tone shifts, hesitation, or drops in empathy—leaving exchanges flat and transactional. Our solution is Glowy, a luminous AI copilot that fuses real-time signals like sentiment, tone, and facial cues into a single “confidence pulse.” Glowy surfaces coaching moments exactly when clarity or warmth dips, helping teams respond with empathy and precision. With playful, intuitive dashboards and customizable “constellations” of KPIs, GlowingStar.AI makes complex signals feel approachable and actionable.

searching for & rating AI agents

Code Guardian is a VS Code extension that monitors your code in real time, detects vulnerabilities, suggests fixes, and applies them when you request all using GPT-5. It’s language-agnostic and highly configurable, so you stay in control. With the rise of coding assistants like Codex, security has never been more important. Code is moving from shipping in weeks to days, and that speed demands protection. Code Guardian is a VS Code extension that monitors your code in real time, detects vulnerabilities, suggests fixes, and applies them when you request all using GPT-5. It’s language-agnostic and highly configurable, so you stay in control. Imagine you accidentally commit a hardcoded password. Code Guardian instantly flags it, explains why it’s a risk, and can generate a secure fix you can apply with one click. Security remediation becomes as simple as accepting a code suggestion—keeping your code fast, safe, and production-ready.

"WDYM" - Agentic memory plus augmentation that lets AI understand what you really want. WDYM remembers your personalized preferences based on previous chats, and augments your prompts to say "what you really mean" - so you can get a personalized experience with faster and better results across every AI Interaction.

Give the average consumer the ability to deploy swarms of agents to save you money on your finances through finding and negotiating better insurance policies, loan rates, etc.

BYOB personalizes learning

“Most of us get nervous about singing in front of people—whether it’s karaoke night, the school recital, or even my own performances. I’ve felt that same anxiety, and it inspired me to build Vocalysis: an AI-assisted vocal coach that gives you objective, real-time feedback on your voice. Vocalysis focuses on the skills singers struggle with most: tuning, breath support, range, expressivity, and confidence. You upload or record a short clip, and our listener agent extracts pitch, energy, vibrato, and timbre features mapped to meaningful traits. From there, the interpreter flags what’s working and what needs attention, while the mentor layer translates it into quick drills and encouragement tailored to your practice goal. The effect is immediate. Instead of guessing why a note felt shaky or a phrase fell flat,

We are trying to create a simpler user surface area for finance and accounting professionals to automate excel wrangling. This tool allows them to create python scripts that run in sandboxes on their information, and provides them a way to evaluate the robustness of those scripts to gain confidence in them, without having to touch the code

Self-enriching, self-improving database for business intelligence. Sometimes, when you're asking questions of your data, there are gaps in what values are available, or sometimes you're not even tracking the data in the first place. I built a multi-agent business intelligence system that does three things: 1. Answers questions about your database from the information being tracked 2. If there are missing pieces of information, it'll go out and search for it on the internet 3. If it turns out the database isn't even set up to track the information in the first place, then it'll suggest changes to the schema to start tracking that information for all entries What you end up with is a database that when you ask questions enriches itself if it can't answer you and improves itself if it's not set up to answer you.

all about AI undertsanding people

Claryn is an agentic automated ML as a service platform that builds and improves machine-learning models for teams needing instant, auditable outputs without hand-coding pipelines as well as people who want to build prediction models without formal training; it automates data profiling, model training, fairness checks, and artifact management while explaining its reasoning. Users can upload a dataset to have Claryn evaluate multiple algorithms and return a trained model, notebook, and metrics, or supply a baseline model plus new data to trigger automatic retraining with promotion/blocking guidance and refreshed reports. The platform also bundles tooling like a Database Hub, Auto-Train Studio, Model Improvement workflow, live runs/logs, and a model registry to manage datasets, orchestrations, and artifacts end to end.

AI clone with lip sync realtime - web search, computer use and sentiment analysis for loading state expression cloning

Soultrace is an AI-powered application that bridges ancient Eastern wisdom with personal digital identity. It addresses the modern problem of fragmented self-understanding by providing deeply personalized, philosophical insights. Users input a three-digit number, which the app interprets using the I Ching (The Book of Changes) to generate a personalized hexagram. They also share cities meaningful to them. Soultrace then synthesizes this hexagram with their current location, generating an AI-powered life analysis. Leveraging GPT-5's web search capability, the app identifies key city landmarks. It then creates a unique "emotional melody" for each city based on the user's memories and identity there. This melody informs the generation of high-quality, emotionally resonant AI images of those landmarks. Finally, all elements are woven into an original poem and a stunning video, merging poetic elegance with visual artistry for a profound, introspective experience.

One of the biggest problems in today’s AI-driven recruitment is credibility and transparency, as a result of which the process efficiency also takes a hit. So, I am working on an MVP - which attempts to make hiring more credible, transferable and transparent for both candidates and recruiters.

1ClickDivorce is a web-based platform that simplifies and streamlines the separation process. With guided questionnaires and AI-powered recommendations, it connects you to the right professionals—lawyers, financial advisors, and counselors—based on your unique situation. The platform offers clear, cost-efficient, and unbundled services, ensuring transparency at every step. Designed to be unbiased and gender-neutral, 1ClickDivorce empowers individuals to navigate divorce with clarity, confidence, and compassion.

Build a real-time, multi-agent courtroom simulator powered by GPT-5-Codex that allows users to present legal arguments and receive dynamic feedback from AI agents playing the roles of Judge, Opposing Counsel, and Jury. The system predicts trial outcomes and highlights persuasive strategies.

Cursor for 2d Video generation

StoryFactory uses voice input to capture an outline brief and then a mesh of agents work together to turn it into a detailed brief

In disaster zones, every minute matters. Drones can deliver supplies and locate survivors, but today responders must manually program waypoints or rely on technical operators -- a slow process when lives are on the line. AeroAid solves this by supercharging the industry-standard PX4 autopilot and MAVLink protocols with GPT-5. Instead of coding waypoints, responders simply give high-level instructions like: “Survey the flooded zone and deliver supplies to the school gym.” AeroAid interprets the intent, decomposes it into subtasks, and generates validated PX4/MAVLink missions that drone fleets can execute immediately. Our system integrates natural language processing, mission validation, and multi-drone coordination into a single pipeline. In simulation, drones receive GPT-5 generated missions, execute them, and stream telemetry (position, battery, wind, speed) back to a central dashboard for situational awareness. We demo using Gazebo Sim.

adds the /agents functionality to codex-cli tool

Miniapp fully onchain game

Cogniforge is an adaptive reinforcement learning environment that lets anyone program robots with plain English. Industrial robots today are brittle: even a 2 cm misplacement can cause failure, and programming new tasks takes weeks of expert engineering. Cogniforge solves this by combining GPT-5 and Codex with a custom PyBullet RL environment. A user gives a natural language command, and Cogniforge automatically generates a behavior tree, an expert demonstration, a rapid imitation learning policy, and a reinforcement learning fine-tune. Before acting, the robot uses GPT-5’s multimodal vision to detect real-world offsets and adapt on the fly. The result is a robot that learns and corrects itself in seconds — turning weeks of robotic programming into minutes.

Spawn the backbone of every software, live tools; apis, bots. cron jobs an beyond

DoorDash for dining halls so students don’t need to leave their dorms when they need to lock in.

this is a Semantic Memory Chrome extension that creates a local-first browsing memory system powered by GPT-5. It automatically captures and analyzes your web browsing activity (page content, highlights, form interactions, scroll depth) and stores it locally with vector embeddings for semantic search.

Paperclip is an app that turns any long-form content (e.g., research paper, Substack blog post) into a series of short-form videos designed to be educational and engaging at the same time. The feed of the app is populated by short-form video from the content the user uploads. If the user scrolls vertically, they'll see new videos. But if they scroll horizontally, they'll see another clip from the same article or content. The problem I'm trying to solve is mainly social media consumption where people go on Instagram and TikTok not to just doom scroll but to seek inspiration at the same time. However, we don't have control over the algorithm. This app allows people to design what they want to consume and the AI feature makes those contents more engaging. We want to give back control to the users, but also help people learn better and gather inspiration by transforming dense content into more engaging formats.

Building Dexter, a secure and searchable knowledge for consultants. From our consulting experience, we saw that consultants cannot easily share or upload slide information from cases because of confidentiality concerns (i.e. slides contain the client name and revenue financials). As a result, it’s hard for people to store information in an internal database and query it to learn more information about different case work. Our solution is a web platform that lets the user upload slides and uses GPT-5 to apply redaction rules (i.e. filtering out client names, financials, and rewriting information to make it less identifiable). We use C# to parse the underlying XML of the slides, then pass the information to the python layer for the AI processing step. Specifically, we use GPT-5 to analyze the text as well as images. Then we store the information in Supabase, which we access later when the user makes a query.