Gemini 3 Paris Hackathon

AdSuperSense: The Next Generation of YouTube AdSense Problem: Standard mid-roll ads break video flow, causing viewer frustration and churn. Solution: We eliminate jarring interruptions by blending ads seamlessly into natural transitions using Nano Banana 2 and Google Veo. How It Works: - Smart Hotspots: Our algorithm identifies natural sequence breaks. Creators select a hotspot and assign an advertiser. - Frame Extraction: We isolate the final frame of the current scene and the first of the next. - Product Integration: Nano Banana 2 (Gemini 3 Flash Image) composites the product perfectly into the frame. - AI Video & Audio: Google Veo generates smooth video bridging the scenes with the product composition, plus native audio for the ad script that matches the scene's ambiance. Result: Contextual, on-demand ads that preserve immersion, maintain flow, and maximize advertiser value.

ShareLock is an AI-assisted reconstruction tool that helps turn fragmented witness memories into visual leads. Witnesses describe what they remember in natural language, and Gemini converts these descriptions, along with optional blurry images, into a structured facial representation that can generate candidate reconstructions. The system allows users to adjust confidence levels, iteratively refine features, and merge multiple independent witness accounts into a consensus composite. ShareLock helps investigators, journalists, and search teams quickly produce visual reconstructions when only partial or uncertain information is available.

**VibeCheck** — a UI/UX research agent that automatically audits any product's onboarding experience. You paste a URL. The agent uses Playwright to navigate the signup flow step by step — landing page → primary CTA click → auth/sign-up screen — capturing a full-page screenshot at each step. It then sends those screenshots to Gemini 3.1 Pro, which returns a structured design analysis: the product's design language (colors, typography, UI components, animation vibe), a mapped user flow with decision points, inferred design decisions, and vibe keywords. The result is a visual storyboard of the onboarding journey — each captured screen shown in sequence with Gemini's analysis — so designers can instantly see how a product presents itself to new users, without manually clicking through dozens of flows themselves. **Stack:** Next.js 15 (App Router) · Playwright · Gemini 3.1 Pro Preview (`@google/genai`) · TypeScript · Tailwind CSS 4

next-state turns real-world videos into live simulations. when a person enters a room, they do not just see who is there — they instinctively infer who people are and what might happen next. ai is getting good at scene understanding, but it still struggles to anticipate near-term real-world dynamics. we call that missing layer situational intuition. my project is an attempt to give ai that intuition.

StateShift efficiently finds semantic events in long videos using Gemini and classic search algorithms. A user uploads a video and asks a natural-language query like “Find when the bottle gets knocked over” or “Find when the door opens.” Instead of scanning every frame, the system first compares the semantic state at the start and end of the video. If they differ, it performs a binary search to localize the transition. If not, it uses exponential search (1, 2, 4, 8, …) to discover a likely bracket before refining with binary search. In the common single-transition case, this reduces Gemini calls from linear to near-logarithmic growth, often cutting model usage by 99%+ on long videos. This reduction makes using Gemini for these sorts of tasks a complete no-brainer, rather than a potentially expensive solution. It also expands the possible videos that can be analyzed beyond what can be handled by simple CV models. Two demo videos are included; see the README for setup and usage.

Our gradiospace helps redesign a room from a simple photo. The AI detects the main pieces of furniture, reconstructs a simplified representation of the space, and then allows users to test a new layout before generating a visual rendering.

DeepDiligence is a multi-agent VC due-diligence assistant. It ingests a startup’s PDF deck, website, GitHub, and YouTube demo, then runs four Gemini 3.1-powered agents: ingestion, tech audit, market audit, and GP synthesizer. The system streams its reasoning via SSE (“thinking mode”), produces Mermaid and PlantUML architecture visuals, and returns a structured Google AI Fund Fit Score plus GO/WATCH/NO GO verdict with defensibility, AI integration, market size, traction, and ethics checks.

GeminiZ is an open-source privacy governance layer for AI agents, built as a fork of Gemini CLI. It uses a declarative file called PIMS.md (like .gitignore but for your personal data) to control what PII each tool and skill can access. PII detection runs locally using open-source models (AI4Privacy, Gemma) so your data never leaves your machine. Tools must declare what personal information they need. No declaration, no data. Every access is logged with a full audit trail, giving developers compliance with GDPR and the EU AI Act out of the box. With 40,000 exposed OpenClaw instances and 400 malicious plugins in the wild, nobody had built the privacy layer for AI agents. GeminiZ is that layer.

Stocks get mispriced all the time, not because markets are irrational, but because most investors are looking at the wrong numbers. Accruals, momentum, asset growth, profitability. These signals have been in academic journals since the 90s. Nobody made them accessible. We did. Market Blind Spot lets you type any ticker and get a Blind Spot Score built from six peer-reviewed financial signals, plus a Gemini-powered explanation of what the market is likely missing. No jargon. Just a straight answer. ValuationQuest is the game built alongside it. It teaches you the models behind those signals, Comps, DCF, LBO, M&A, one decision card at a time. Get it right, Alex celebrates. Get it wrong, Morgan the MD tells you what a real banker would think. One day. One stack. Python, Flask, React, and Gemini 3. Thirty years of academic finance research, accessible to anyone with a browser.

CommuniSense is a community-adaptive moderation layer built on Gemini. Most moderation systems assume language norms are universal. In reality, communities speak differently - different dialects, slang, reclaimed language, and cultural context all affect how messages are interpreted. Current toxicity detectors do not account for that and now people are getting **banned for speaking their dialect**, while harmful language specific to a community goes undetected. CommuniSense uses a Gemini-powered agent to study a community, it's rules and discussions, and infer a moderation profile. Instead of relying on static keyword filters, it predicts how members of that community would react to a message and combines this with broader safety rules. The result is context-aware moderation that understands your community.

Live Demo : https://ripple-engine.vercel.app/ The Ripple Engine: A Sovereign Causal Radar for Urban Disruptions. Standard navigation apps show operators the present (e.g., a closed transit station). The Ripple Engine uses AI to map the "Invisible Strings" of the city, showing operators the future. Built for high-stakes sectors like Government Policy, Grid Control, and Fleet Operations, our engine moves beyond generic chatbots. We use Gemini to pre-compute the socio-economic and spatial ramifications of major urban failures (like a catastrophic RER B failure or a mass egress from Stade de France). By applying specific "Profession Lenses" to these signals, the engine generates deterministic, zero-latency directives (e.g., [DIVERT], [SHED], [STAGE]) and visually maps the resulting risk zones. It is a predictive instrument for operators who manage millions of euros and public safety.

RECUT is an AI-powered video genre transformer that takes any ordinary video and re-imagines it as a cinematic production in a genre of your choice — Horror, Action, Romance, Comedy, Thriller, or Epic Documentary. The pipeline uses CLIP (ViT-B-32) for local semantic frame analysis and scoring, Gemini 2.5 Flash as an AI director that crafts a narrative timeline with voice-over scripts, Edge-TTS for cinematic narration, and MoviePy for rendering with genre-specific visual effects (color grading, vignettes, film grain, Ken Burns zoom). It also features a "Moment Finder" — a semantic video search engine where you can type natural language queries like "red car passing" or "someone smiling" and instantly find the exact moment in any video using CLIP embeddings. The app solves the problem of making cinematic content creation accessible to anyone — no editing skills needed, just upload a video, pick a genre, and get a fully narrated short film in under 90 seconds.

Our project's goal is to generate customized sports highlights from video of events. For example, our user would love to get highlights of its favorite player: Luka Doncic, that's why he needs to complete information about the game, upload the video of the game and then it generates a video that combines all actions of Doncic during the game. Our project could be in the future extended to multiple sports (football, rugby, american football, cricket).

Seedling is a multi-tribe civilization simulator powered by AI agents. It explores a central question: "How does a society's willingness to change its own rules shape its long-term trajectory?" There are four tribes being simulated: The Keepers, The Moderates, The Adapters, and The Free. Each start with distinct governance rules and different levels of rule mutability. Every agent is an autonomous LLM-powered individual with its own personality, values, memories, and relationships, reconstructed from database state before each interaction. Agents deliberate in shared-thread round-robin councils, propose and vote on rule changes, form partnerships, and respond to random events (famine, discovery, conflict). Cultural values drift across generations as children inherit blended values from parents with mutations. The simulation compresses years of societal evolution into minutes, streaming results in real-time via SSE.

Monorepo layout for parallel MVP delivery of a French public-interest assistant that turns official open data into clear, sourced answers for citizens and public agents. additional video: https://wetransfer.com/previews/14b6b039844fc52fc4aca821d953404220260314155542/477e20?itemId=105dc84f56cace1b39165b7240ac12a220260314155650

Great founder stories are buried in long interviews, blog posts, and PDFs scattered across the web. FounderPod helps you learn about any founder story in a fun format and within just a few minutes.

Vibly is an AI-powered digital wardrobe web site and fashion advisor Chrome extension that allows users to virtually try on clothing from any online store. The AI analyzes the fit, style, and compatibility with the user’s body and wardrobe, providing personalized styling feedback to support purchase decisions. It also generates outfit recommendations by combining items from the user’s existing wardrobe and suggesting looks tailored to specific social occasions. It can also automatically read user's Google Calendar to get social event of the day and make recommendations https://ais-pre-x3szvzpy25haawjk3u6op7-479442645157.europe-west2.run.app

agorAI is a multi-agent negotiation marketplace powered by Google Gemini. Users create autonomous buyer and seller agents with various negotiation styles and characteristics. When a buyer posts a request, the platform discovers relevant sellers via semantic matching, then launches parallel Agent-to-Agent negotiations. A central coordinator evaluates all resulting offers on price, quality, and flexibility, and selects the best deal, fully autonomously.

different thinking system different image generation

This project transforms a standard Unitree robot dog into an intelligent, visually-aware companion powered by the Gemini Live API. By connecting the robot to its multimodal Gemini-based brain, it easily understands its environment and allows users to ask visual questions and perform actions.

I work in climate policy. Most days bring bad news: missed targets, stalled legislation, another funding cut. After a while, the setbacks become numbing. Then I saw a short video asking: what if it all goes right? I couldn’t stop thinking about it. So I built Positive What If, a time machine for optimists. Users choose a future they care about: universal basic income, staying under 1.5 °C, or cities designed for people, not cars. The app loads ~50k tokens of curated research and generates a narrative set in a real city and decade: what worked, what failed, the trade-offs, and feasibility scores across four dimensions. Every claim is cited. It also generates before-and-after street images and a shareable postcard from that future. I built it because I was tired of imagining only the worst.

https://handshake-iq.replit.app/ Every IRL meeting leaves a dead handshake. It happened. It mattered. It disappeared. No CRM captures it. No AI can record it. Handshake IQ fixes this. Drop scribbled notes or a photo of handwriting into the system. Gemini 3.1 reasons across it and logs a relationship event that builds a compounding score — like a credit score for trust. Live camera scores the handshake itself. Grip. Eye contact. Posture. Duration. We call it a Handfy. Not CRM. Infrastructure. The kind Bloomberg built for markets. Today: insurance. Tomorrow: a country relationship score every time world leaders shake hands. IRL meeting is human intelligence at its best.

GenealogyAI autonomously traces French family trees by navigating historical civil registry archives, given a name, a town, and a birth year. The problem it solves: tracing French ancestry today requires hours of manual work, and it requires searching fragmented departmental archive websites one by one, identifying the right register, opening scanned images, and reading 19th-century handwriting. No existing tool (MyHeritage, FamilySearch, Geneanet) does the full loop: autonomous navigation of raw archive systems + OCR on unindexed manuscript images + recursive extraction. Our tool does all three in a single query.

PermisEasy is a hands-free driving lesson companion. Launch the app, put your phone down, drive. It silently records your GPS route and conversation with the instructor. After the lesson, Gemini 3 analyzes the audio to detect every mistake and advice, then maps them onto a color-coded route. Click any error to see the intersection in Street View with an AI explanation of the traffic rule you broke. In France, 50% fail their driving test — not because of technique, but because they can't retain road rules in real situations. PermisEasy makes every lesson reviewable, visual, and AI-enhanced.

● CYNIC is an epistemic immune system — 4 independent AI models (Gemini 3, Llama 3.1, Gemma 3 sovereign, deterministic engine) evaluate any content through 6 philosophical axioms, reaching consensus under mathematical doubt. Confidence structurally capped at 61.8% (golden ratio inverse). Demo: real-time chess judgment — each move evaluated by 4 AI validators through philosophical lenses. When AIs disagree, it's a discovery signal. The system learns through Cognitive Crystallization: verdicts accumulate into persistent wisdom. Built today: Gemini 3 integration, multi-model consensus, per-Dog latency tracking, CCM learning loop, Dog selection API (33ms to 13s), and React chess dashboard. 17 commits, 8000+ lines on our open-source Rust kernel.

Disrupt personal brand saturation. Map irrelevancies and semantic resonance to synthesize your hyper-individualized lifestyle and fashion archetype.

ARCADIA brings back the magic of '80s arcades, where you walked in blind and discovered new games. No spoilers, no trailers, just unknown bangers tailored to your vibe, unlike recent videogames. Here's how it works: Black screen. AI (Gemini 3) chats your hobbies/mood → explodes a never-before-seen 2D game mixing '80s arcade DNA with your personal twists. Finish? Poof—deleted forever (FOMO maxed). Single-player hackathon MVP by our 3-person team. Tech: Gemini 3 multimodal → procedural WebGL chaos. Demo live: [INSERT LINK] Future enhancements: MultiPlayer matchmaking + voice chat with vibe-matched users.

I work in Cinema and every artist i know feels (1) threatened by AI and (2) neglected by current AI projects/apps/solutions. Besides, nearly all film productions or sets I've been in run on the same broken workflow: a PDF screenplay, shared and updated daily via a 100 person WhatsApp group. My project solves this by creating an entire platform to organize all artists using the screenplay as the single source of truth. Uploading a PDF script creates a structured database with every scene, character, location, and line. From that one upload: live call sheet, shooting schedule, weather forecasts, and natural-language interface to organize teams or ask anything.

Vibe Research engine

Lumen-Sonic turns any light source into an original music track. Point your phone at a lamp, candle, or neon sign — Gemini 3 Vision analyzes the light's color temperature, flicker rate, and brightness arc, then writes a music prompt. Lyria generates a 30-second WAV file from that prompt. No two lights sound the same, and the same light can sound different on different phones due to sensor and white-balance variance.

Earth-Mars latency makes direct control impossible. You can't pilot the robot — you can only give it orders and wait. The game is in the prompt: vague orders fail, precise orders succeed. Mars Prompt Arena is a competitive training ground for finding the world's best remote robot operators. Players take on the role of Mission Control on Earth, sending natural-language commands to CANIS-1 — an autonomous robot dog stranded on the Martian surface. There's no joystick. No WASD. Just your words, an AI brain, and a ticking clock. The robot runs on Gemini — it reads your prompts, plans its actions, executes them in a physics simulation, and reports back in first person like an astronaut on comms. Your job is to phrase commands that make it do the right thing.

ITER-ATION is a real-time monitoring system for nuclear fusion, inspired by the parameters of the ITER experimental reactor, in which a Gemini-powered AI agent continuously analyzes plasma data and takes logical corrective actions—such as gas injection or power reduction—to maintain tokamak stability and prevent disruptions.

Oppy — AI Voice Assistant for Multi-Project Management Oppy is a voice-first AI assistant built on Google's Gemini ecosystem that helps students and professionals juggling multiple projects (school, work, startup) stay on top of everything. It connects to Gmail and Google Calendar, scores urgency across all projects, and delivers spoken briefings through natural voice conversation powered by Gemini 3 Flash (chat) and Gemini 2.5 Flash TTS (voice). You talk to Oppy like a mentor — it tells you what's urgent, who's waiting, and what to do next. Key features: wake word activation, persistent multi-turn chat, animated SVG avatar with eye tracking, real Gmail/Calendar integration, smart acknowledgment during loading, and a Gemini-inspired UI. Unlike generic chatbots, Oppy is opinionated — it doesn't ask, it tells. "Sophie hasn't heard from you in 6 days, reply tonight." Voice-first means you get briefed while cooking or commuting.

Project Name: The Passenger Pet Tagline: Zero-touch intake and logistics for animal rescue.

Seamless futuristic task manager for normal people

Radioshaq was inspired by living close to the french / spanish border and mountains. based on my experience , radio operations need this , it's that simple. RadioShaq brings citizen radio operations into the 21st century, it's the first and only opensource system and citizen system that does simple things for example multiband relay , message store and send , etc. everything is compliant by default , whitelist based , gemini powered. So what does it actually do ? the gemini-powered orchestrator is supported by a judge and deploys subagents and subjudges for each task . for exa, mple "radio chat" , for example "relay" , for example "store location information", and respond with coordination and other compliance according to stored and persistent data.

We believe there's a gap between personal context and agentic workflows. AI agents can execute tasks, but they don't know how you specifically do your job - which apps you use, in what order, what rules live in your head, what exceptions matter. That knowledge is trapped in muscle memory. No agent can read it, no automation tool asks for it, and there's no structured format to store it. Jemo bridges that gap. You share your screen, do the task once, and narrate your logic out loud. The app watches and listens, then builds a visual workflow graph in real time - turning personal, unspoken process knowledge into a structured, machine-readable format that agents can actually act on. Think of it as a personal context layer for AI. Not another automation builder — a way to teach an agent how you work by simply showing it.

Croissant Toolkit is an intelligent suite of tools designed to work seamlessly with the MLCommons Croissant metadata format for machine learning datasets. Powered by the incredible capabilities of Google Gemini 3, this toolkit simplifies, validates, and enhances ML dataset preparation, discovery, and exploration.

CoEvoLve — A Living Benchmark That Evolves Faster Than Any Model Can Game It. Static benchmarks leak. CoEvoLve solves this with a coevolutionary system where the benchmark and the solvers evolve together in an open-ended arms race.Two populations of Gemini agents co-evolve simultaneously: Examiners generate causal reasoning tasks and evolve toward maximum discriminative power — tasks that separate strong solvers from weak ones. Solvers attempt those tasks and evolve better reasoning strategies. .After 500 generations, the system independently discovered that graphs combining confounders, mediators, and colliders are maximally discriminative — a finding that matches the theoretical causal inference literature, without being programmed. The benchmark found the structure of LLM failure automatically.Built on ShinkaEvolve (Sakana AI, Apache 2.0) extended with dual-archive coevolution, Gemini 3 Pro for examiner mutations, Gemini 3 Flash for solver evaluation.

Project that allows to see Gemini interact in a world generated by AI gemini too.

Our platform brings the next generation of virtual artists to life by instantly transforming visual inspiration via Gemini Vision into a live instrumental with Lyria RealTime, seamlessly fused with a unique vocal signature to generate the complete branding and music video for a new AI popstar. https://vibemusic-eosin.vercel.app/

FabricMind - AI-powered digital twin system that visualizes, monitors, and optimizes fabric production workflows in real-time with interactive 2D isometric representations.

Thread is an AI-powered external working memory that helps people recover their train of thought after interruptions, multitasking, or cognitive overload. Modern life constantly fragments our attention through notifications, meetings, and context switching. When we return to a task, the hardest part is often remembering what we were doing and why. Thread lets users capture fragments of thinking—text notes, screenshots, or voice input—and uses Gemini’s multimodal reasoning to reconstruct a structured “thought thread.” The app highlights what you were working on, your goal, the last step you took, and clear next actions. Inspired by Ariadne’s thread in Greek mythology, Thread helps users find their way back through the labyrinth of interruptions while supporting cognitive accessibility and reducing mental effort.