Gemini 3 Hackathon SF

NNPort redefines embedded engineering by fully automating the tasks that traditionally required specialized engineers, delivering an end-to-end autonomous development workflow. Instead of replacing engineers outright, NNPort empowers them with advanced vibe debugging, step-by-step code exploration, and transparent logic tracing. This allows developers to understand, refine, and optimize individual code blocks with greater clarity, speed, and confidence.

ForenSight AI is a multimodal “AI detective” that lets you upload scene photos, videos, audio clips, and witness statements - and instantly generates a unified, AI-reasoned reconstruction of what likely happened. Gemini 3 analyzes every piece of evidence, builds timelines, finds contradictions, and proposes multiple plausible scenarios, while Nano Banana Pro generates high-resolution 2K/4K visual reconstructions from those scenarios. The goal was to explore how far multi-modelity can go when reasoning (Gemini 3) and rendering (Nano Banana Pro) are fused into one interactive forensic experience.

Chess Commentary AI is a real-time AI-powered chess commentator that uses Google Gemini Flash to watch your screen and provide exciting play-by-play commentary. It captures your chess game via screen sharing, analyzes the board state using AI vision, and streams live text commentary.

Anywhere is a Next.js/React app that turns Google Street View into an AI-guided, voice-controlled tour guide. It solves the “WebGL Street View has no clickable UI” problem by letting Gemini issue structured navigation functions (pan, move, teleport) instead of brittle browser automation, and adds live Google Search grounding for accurate facts while you explore. Users speak to navigate, get context-aware narration, and can generate AI “selfie” composites of themselves in the current scene.

Clear Witness is an AI-powered forensic reconstruction suite designed to aid law enforcement and investigators in regions where digital surveillance (CCTV) is scarce. It democratizes the capability of a professional forensic artist, allowing users to reconstruct high-fidelity suspect imagery using only eyewitness descriptions or rough sketches. The application features a "Character Lab" where users fine-tune physical attributes (DNA, bone structure, aging) using granular sliders. It includes a "Digital Sketchpad" equipped with professional tools (Mesh Warp, Liquify/Pinch/Bulge) to manually correct features before sending them back to the AI for photorealistic synthesis. A key innovation is the "3D Turnaround" generator, which uses Gemini to create a multi-angle character sheet that the UI renders as an interactive, rotating 3D view, providing investigators with a holistic view of the suspect. All generation and processing happen securely in the browser using the user's own API key.

FlowState is the first visual linter for frontend code—a tool that bridges the gap between design and development by letting you literally point at broken pixels and watch them heal.

intelligent assistant to help students use and control their device.

Welcome to the world's first AI generated sports betting network. The sports are generated by Gemini, visualized by Nano Banana and a play-by-play is delivered, Veo will on demand create short clips that act like highlights you'd see on ESPN and correspond to the play-by-play and the image/keyframe Nano Banana generates - creating a continuous virtuous prompt-to-entertainment loop. You can watch and get play by plays of entire 'games' this way, on demand as a long running process. You can also, importantly, bet on the sports. In the future multiple people will bet, and the network will continuously run the sports so they happen in sync for a fair, random experience for all - and it'll cache the veo videos for faster loading rather than current behavior, on the fly

"SpecForge" is an Instant Product Factory that solves the engineering "blank page" problem. It transforms a simple text description into a comprehensive engineering design packet in under 60 seconds. It automatically generates rigorous JSON specifications, technical blueprints, a Bill of Materials, firmware/application code, and even a cinematic product commercial. It allows rapid prototyping by bridging the gap between a vague idea and a validated, buildable design.

AJs Animator transforms standard "talking head" videos into engaging, multi-sensory educational content. The tool analyzes speech in real-time to automatically generate and overlay dynamic, context-aware visuals, helping creators explain complex concepts and boost viewer retention without manual editing.

A Google VISTA-inspired prompt optimization framework for text-to-image generation using SigLIP2-based quality scoring and DSPy's GEPA optimizer.

Darkroom is a multimodal AI application that restores, enhances, and animates historical photographs using three Gemini models in a unified workflow. Gemini first detects photo boundaries and extracts historical metadata including era, location, clothing styles, and socioeconomic context. Gemini then restores damaged photos, removing scratches and fading while optionally colorizing black & white images with historically accurate tones. Veo 3.1 (Nano Banana Pro) generates cinematic videos that bring restored photos to life with subtle, natural motion. The app automatically fetches related Wikipedia content to provide historical context and teach users about the era, clothing, and lifestyle depicted in each photo. :)

This is a test submission.

- VibeCheck is a "vibe debugger" that will review a PR and compare the implementation against the Requirements given in a Requirement description and then review the PR to see if it meets the requirements or not. It will also debug and suggest fixes to the PR so that it matches the requirements. - VibeCheck can also add unit tests to the PR code as well. - In addition to code verification and Unit Tests, VibeCheck can also compare UI design elements and compare the implemented UI against the UI design from the requirements by comparing the screenshots of the implementation vs the requirement and commenting on the visual differences.

Your photography PAL. Minimize time spent on editing and selection during professional photography by using PAL to automatically: - Categorize different image shots into buckets - Semantically rank them by comparing them to other relavent images - Running custom post-processing filters on it using Nano Banana This saves hours of time, and can work in several industries: - Wedding Photography - Casual Photography - etc....

This project is a Multi-Modal Code Refactorer with Visual Verification, built in one day. It solves a common and time-consuming problem in software development: understanding and improving messy or legacy code while ensuring that UI components look and function as intended. The system allows a user to upload backend or frontend code, including React components, HTML, and CSS. It then uses Gemini 3 and Nano Banana Pro to refactor the code, improve structure, apply best practices, and explain its changes. For frontend code, the system also runs the code in an isolated environment and captures a real, non-hallucinated rendering of the UI. Gemini then reviews the real visual output and provides a corrected or improved version of both the code and the interface. This provides a fast, accurate way to clean up codebases while ensuring the UI behaves as expected.

All in one brand asset management tool - Manage your brand's identity (mission, core aspects, color pallete) - Create and iteratively modify asset designs (hoodies, billboards, flyers) - Brainstorm and create full fledged advertisements with voiceovers Brand visibility matters! and we help you get there

AI-powered documentation generator that not only writes API docs but verifies every code example actually works. Option of other languages available.

Spins turns any prompt into a dynamic, AI‑generated audio experience you can steer in real time. Stories, briefings, explainers, etc. as playlists you can reorder, remix, and expand as they play.

At Chlora-Flora, we catch diseases at one plant instead of 1000. Farmers analyze areas of crop fields while we analyze individual plants and provide data for plant growth and yield. We're doing what Palantir and Polymarket did, but for food.

System to launch many sandboxed coding agents with code execution and testing.

RIddler! Make quizzes ASAP

SmartFetch is a persistent stateful agent that fixes client API failures and prevents outages across all services, regions, and providers. It detects API key deprecations, rate limits, quota exhaustion, schema drift, endpoint changes, regional failures, and provider outages in real time. The system repairs malformed requests, infers new schemas, rotates credentials, selects alternate endpoints, and switches to healthy regions or providers when services fail. Gemini provides the reasoning needed to classify errors, generate corrected payloads, and choose the right healing action. The system learns new patterns over time and builds custom healing rules for each API. This creates a fully autonomous reliability layer that solves a broad range of client-side API problems and keeps applications functioning during failures of any kind. In the video we tested fixing a broken API connection to OpenAI, which is quite funny.... Gemini fixing the mistakes of OpenAI.

VoiceAI to compare performance of openAI, gemini, Claude models

As a business owner, I'm constantly making new creatives. I built a tool to help me iterate on my ads and build the assets I need quickly. This helps me create new ads from scratch or change existing ads to promote a sale or change to a new audience. "Good artists copy, great artists steal" definitely applies, use working ads to make new ones!

Most AI tools give you answers, but Gemini can be a teacher—one that adapts to how you think. Dual Class proves this by translating the same technical concept into metaphors that represent your profession, hobby, or passion (a chef's kitchen, a starship bridge, a detective's case board) and using visual evaluations to confirm comprehension. It's adaptive learning that matches your existing expertise and background knowledge.

Clueless: The Multi-Modal Stylist Agent that shops your own closet. We all face "wardrobe paralysis"—staring at a full closet with "nothing to wear." Existing solutions are just static catalogs or basic body-shape classifiers. They lack the imagination to visualize how items fit together on you for specific occasions. Clueless is a generative stylist agent that solves this by bridging vision and creation. We use the Nano Banana Pro model to generate photorealistic "Try-On" visualizations, rendering the user wearing the suggested outfit with correct physics, lighting, and fit. Clueless creates a rich multi-modal loop: Input (Text + User Image + Wardrobe Images) → Reasoning (Gemini) → Visual Synthesis (Nano Banana Pro). It moves beyond text advice to visual proof. Tech Stack: Backend: NextJs + Google Gemini SDK Frontend: React + Vite (Tailwind CSS) with "Digital Vogue" aesthetic AI: Gemini 3 Pro (Vision/Reasoning) + Nano Banana Pro (Image Generation)

Built Game Scroll that builds a social media website to allow people to view different games that is a demo version of the real game giving them a feel, allows them to give instant reaction and validation and comments to allow developers to help them fix it.

BananaCutPro is an advanced AI-powered design and image editing tool that empowers users with "Vibe Designing" capabilities. By leveraging the power of Google's Gemini models and Composio, it bridges the gap between high-level creative intent and pixel-perfect execution. Whether you need to make precise edits, batch process complex changes, or get expert design analysis, BananaCutPro provides a seamless, professional-grade workflow similar to tools like Figma but supercharged with AI.

Interium designer which actually can find the products

The Living Playbook: Gemini as a Sales Teacher We focused entirely on Statement Three: Gemini as a Teacher. In high-stakes sales, "getting an answer" isn't enough; reps need to learn how to win. We built a system where Gemini acts as an active Sales Teacher that helps humans rapidly level up. Instead of static PDFs, our system ingests real-world failure cases (audio/video calls) and autonomously creates derivatives (new playbook rules) and feedback loops (retrospective coaching). It transforms unstructured conversation data into a structured curriculum, ensuring that a junior rep learns the deep "domain context" instantly. We turned Gemini into a mentor that reasons through mistakes and updates the lesson plan in real-time.

Using a generalized Sylvester equation and Gemini 3 Pro preview, we're able to bridge the gap between videos and LLMs.

vibeCAD is a browser-native, parametric CAD system built with TypeScript and React, featuring constraint-based design and 3D solid modeling powered by OpenCascade.

GiftGenius solves holiday gift-giving stress by turning a simple text description into a beautiful, personalized gift guide. Users describe the recipient (age, interests, hobbies, budget) and get a high-resolution visual guide with 3-5 tailored gift recommendations, each w/ pricing and personalized reasoning. It's shareable, printable, and perfect for the time-crunched holiday season.

Crack - An agentic system which understands brand DNA and deploys coordinated agents to create hyper-personalized targeted ad campaigns for small businesses. An agentic system that: 1. Extracts brand DNA from a URL - aesthetics, tone, assets, business overview 2. Deploys parallel agents that coordinate to create personalized campaigns: - Orchestrator (main thread) - Planning agent - Segmentation agent - Image generation agent - Font render agent - Stitcher agent 3. Targets intelligently - collects user preferences, then does selective personalization

Imagine having gemini live with you inside your glasses? it is the next wave of the human to AI co-existance where you have the intellgence not on your hand but with you all the time!

Nichijou (Your Everyday Life) hacks the biggest problem in language mastery: irrelevance. Most apps like Duolingo force adults to memorize generic sentences about schools they aren't attending or libraries they'll never visit. These phrases lack emotional weight and vanish from memory. Nichijou flips this model by turning your actual daily life into a bespoke multi-modal experience. You simply journal about your day in English. In seconds, the app transforms that memory into a personalized manga immersion in Japanese. By generating vocabulary, grammar points, and visuals derived specifically from your experiences, Nichijou creates a rich multi-modal loop where you read, see, and hear your own story. You aren't just memorizing words; you are immersing yourself in the language of your own life, making it immediately useful and impossible to forget.

We provide an entire end-to-end agentic pipeline for embedded systems which use FPGAs (extendable to CPLDs, etc). Users with no experience in verilog can upload/design their C code and have it converted to verilog with an agent orchestrated to iterate and improve it via synthesis/simulation of an FPGA with full test generation and LUT optimization to fit code on smaller and more affordable devices.

Generate production-ready ads from Amazon listings. E-commerce is exploding, and sellers need video ads that convert fast. Using Gemini, we transform existing product images into scroll-stopping video ads with one click. Amazon is just the entry point; the possibilities are endless. Imagine a full end-to-end system: generate → upload → A/B test → regenerate automatically until we hit peak conversion. Set it and forget it. It scrapes Amazon using Brightdata.com, passes the data to Gemini 2.0 Flash to create TTS speech scripts and prompts. Loop image generation until the first frame of each scene is effective. Uses Veo 3.1 to generate each scene. Uses ElevenLabs to generate TTS. Leverages Remotion to produce the final video.

This project demonstrates **“Gemini as a Teacher”** by tackling one of the hardest real-world skill problems that exists: teaching attorneys how to conduct elite cross-examinations. Cross-examination is widely considered the most difficult skill in trial advocacy. Even seasoned trial lawyers struggle with it. Unlike many legal tasks that are repeatable or research-based, cross-examination requires a rare convergence of abilities that almost no existing training systems can replicate: - Real-time psychological reading of a witness - Dynamic questioning under pressure - Strategic sequencing with imperfect information - Emotional regulation in adversarial dialogue - Jury persuasion across varying personality profiles - Instant adaptation to objections and judicial rulings Most attorneys practice cross-examination **without realistic feedback loops**. True exposure to real-pressure cross-exams only happens once attorneys step into court — where mistakes carry permanent consequences.

Form, manage, market, and measure your Non-For-Profit organization, Agentically!

Pookieverse 🪽 Ever wanted to play god but, like, a nice one? Pookieverse drops you into a living world where AI creatures called Pookies roam, think, and make their own chaotic decisions. Your job? Be their Guardian Angel. You don't control your Pookie directly, that's the catch. They have their own personalities (some are greedy, some want to save the world from imaginary disasters). You can only whisper guidance and watch them completely ignore it because they're "feeling brave." The magic happens when multiple players join. Watch Pookies form friendships, start petty drama, negotiate trades, or just yell insults like "pestlepuff" at each other. It's multiplayer chaos powered by Gemini Agents. All the sprites and maps are generated with Nano Banana.

Muse doesn’t just fix music discovery—it fills a gap in how we understand ourselves through what we listen to. Streaming platforms organize songs by genre and vibe, but they miss the emotional context of listening: the way we use music for self-discovery, relationship building, emotional processing, conflict resolution, and getting into deep-focus flow. Muse’s emotional profiling can power: Smart playlist curation – “Find me songs for a rainy afternoon drive with my best friend” or “give me calm-but-confident pre-presentation hype music.” Music therapy & emotional processing – Curate tracks that support specific goals like grounding anxiety, processing grief, or boosting motivation. Relationship & social experiences – Generate shared emotional playlists (“our songs”) that reflect how two people feel, not just what they like on paper.

Flux is the DevOps layer for LLMs, providing version control, testing, and observability for AI prompts. It helps you manage prompts like code with proper workflows, testing, and monitoring.

Ability for Gemini 3 to solve long-horizon tasks in Data Engineering DAG Errors.

Today most people meet history as walls of text, slow documentaries, or random TikToks ripped out of context. They never build a felt, character-level connection to Alexander, Cleopatra, or anyone else, even though their lives are basically prestige TV material. History Daily turns that missed opportunity into a product. You pick a historical figure and a major campaign; we turn it into a season of 30-second POV vlog episodes, like “Alexander with a phone on the road to Persepolis.” Each episode is a dramatized but historically grounded moment with a clear emotional beat, generated with Gemini + Veo. After every clip we surface a couple of high-quality “go deeper” links—long-form video and readings—so curiosity has a clear next step. It’s not a textbook; it’s an on-ramp that makes serious history as bingeable as a creator’s vlog series.

I created a digital watermark for nano banana images that anyone can upload and verify their details using the same image. this makes insurance claims machine readable and authentic because of non hackable nature with aes encryption. This project adds a secondary, tamper-resistant watermark to AI-generated images. The watermark encodes a unique identifier that pairs with existing GenAI watermarks, creating a reliable chain of authenticity. Any system—or chatbot—can decode these markers to verify image origin and integrity. The approach supports real-world use cases like insurance claim validation, fraud detection, and verifying the legitimacy of marketing creatives, without altering the visual quality of the images. It’s a lightweight way to bring trust and traceability to images with generative ai. Images without watermark woudl be useless. with images with watermark two things are validated - synthid (using nature paper code) and the custom watermark to make it work.

Ruvio is an AI-powered interior design platform that transforms 2D floor plans into photorealistic 3D room designs in under 15 seconds using Google Gemini's multimodal AI. Users simply upload a floor plan, and our AI automatically analyzes rooms, generates multiple design variations, and allows natural language customization—making professional interior visualization accessible to everyone without design skills or expensive software.

The vibe2vibe evaluates the total code complexity of a given source code repository on several measures (cyclomatic complexity etc.), then uses this to split the resulting application into smaller microservices. To showcase this for a real applicatoin, we created the Banana Scanner (an LLM Brand Presence Test application we created) as an example target.

Feedie is an autonomous social media agent that creates, optimizes, and manages content across multiple platforms with minimal user input. Users upload media or provide simple ideas, and Feedie generates platform-specific posts, captions, hashtags, and scheduling recommendations. The system simulates a full creative workflow, producing variations for Instagram, TikTok, YouTube Shorts, and more. A dashboard displays real-time job progress, rendering updates, and engagement insights generated by the agent. Feedie eliminates manual content creation by acting as a creative partner, strategist, and publishing assistant—all in one streamlined, automated workflow.

SnapLock is a physics-native generative design tool. We use a WebGPU physics engine (Taichi) to constrain AI imagery, ensuring objects obey "Ground Truth" laws like gravity and collision. Uniquely, SnapLock features a "Self-Healing" Engine: if the simulation crashes, our built-in debugger sends the error to the backend where Gemini 3 Pro triages, diagnoses, and auto-patches the code in real-time. We use Banana.dev for serverless GPU inference (ControlNet) and Gemini Nano for local latency optimization, creating a seamless loop between physics, code, and pixels.

Slide Quest is a live “side quest” for presenters: upload your deck, then summon fresh slides mid-talk with a quick voice description. Gemini 3 handles the multimodal magic while Nano Banana Pro beams each new detour in crisp 2K, so every audience question becomes an instant visual story.

Modern manufacturing environments depend on accurate visual and contextual data to diagnose machine failures, assess wear, and build AI models for predictive maintenance. However, factory floor engineers and data/AI engineers operate in completely different worlds: • Field engineers collect images and videos under noisy, harsh conditions • Data scientists require high-quality, well-labeled, diverse datasets • Miscommunication leads to poor datasets, delayed model training, and production downtime There is currently no multi-modal system capable of guiding field engineers in real time while simultaneously generating structured, high-quality datasets for AI teams.

CoPair AI-powered coding interview prep that adapts to your weaknesses. The Problem It Solves Traditional interview prep is often unfocused—you either grind random problems or don't know where you're actually struggling. CoPair addresses this by: 1. Diagnosing Your Weaknesses — Takes a diagnostic quiz covering 17+ DSA topics (Arrays, Two Pointers, Stack, Binary Search, Trees, Graphs, Dynamic Programming, etc.) 2. Providing Targeted Practice — Instead of random problems, it generates Python coding challenges specifically for topics you got wrong 3. Offering Smart Hints — If you're stuck, the AI provides progressive hints that guide you without revealing the code

An automatic RL environment and RL post training tool for agents. The tool takes a set of logs and creates a RL environment which then in turn can be used to RL post train a small model.

An agent which monitors your project for expired TPAs and suggest patches based on latest news feeds

Most sports broadcasts assume you already know everything. If you’re new, a 90-minute game looks like random running. If you’re advanced, generic commentary doesn’t tell you tactics or context. And on YouTube, most games are just raw footage. We built an AI sports commentator and coach that sits on top of any game video. It watches the match, tracks score and events, renders simple on-screen visuals, and generates live-style commentary that adapts to your level – beginner or expert.

NeuroSensei (Brain Waves + Gemini) is a multimodal AI Speaking Coach that analyzes your words, body language, and brainwaves. Public speaking is hard - mirrors can’t judge eye contact, pacing, or internal stress. NeuroSensei combines Computer Vision, Speech Analysis, and EEG signals to deliver holistic, moment-by-moment feedback for interviews, pitches, and presentations. Key Features • Brain-Sync Analytics: Streams EEG data to reveal real-time stress, focus, and distraction, aligned to video timestamps. • Multimodal Analysis: Gemini 2.5 Flash analyzes video frames + audio to detect posture issues, eye contact, filler words, and pacing. • Time Alligned TTS Coaching: Powered by Gemini TTS, NeuroSensei speaks personalized, time-aligned guidance using a natural, low-latency voice.

Parents and kids can co-create, AI-assisted storybooks using natural language as well as drawing skills, and order a physical copy for themselves.

Lifeline is an autonomous crisis-operations marathon agent that maintains live situational awareness during building emergencies and continuously replans evacuations as conditions change. Gemini 3 Pro handles long-horizon reasoning—tracking zones, hazards, capacities, and accessibility routes—while Nano Banana Pro generates updated 2K evacuation maps by editing a base floorplan with hazard overlays, blocked paths, and clear arrows. When an operator reports an event like “Fire in Zone B,” the agent analyzes impact, updates building state, recalculates safe paths, and issues new instructions. Lifeline demonstrates how Gemini 3’s reasoning and multimodal capabilities can support real-world crisis response with speed, transparency, and reliable map updates.

SafeStack analyzes warehouse video feeds with Gemini to detect violations of OSHA's safety policies, allowing operators to find when there are issues with their warehouse operations. We provide detailed analysis, send emails to alert operators, and use nano banana to generate corrected images of what should have happened. Policies used: https://www.osha.gov/sites/default/files/publications/OSHA3885.pdf

Interactive platform to learn ASL