# Gemini-2 Projects at AI Tinkerers

> Canonical HTML: https://aitinkerers.org/technologies/gemini-2
> Markdown URL: https://aitinkerers.org/technologies/gemini-2.md
> Technology record last updated: 2026-09-18T14:19:39Z
> Generated: 2026-09-22T13:38:59Z

Google’s multimodal model family built for native real-time reasoning across text, vision, and audio.

Gemini 2.0 Flash delivers high-speed intelligence with a 1M token context window: it handles live video streams and complex spatial reasoning (like identifying objects via camera feed) with sub-second latency. This multimodal-first architecture enables developers to build agentic workflows through Google AI Studio or Vertex AI. By processing multiple inputs simultaneously, the model bridges the gap between digital reasoning and physical-world interaction for production-ready applications.

- Official technology site: https://deepmind.google/technologies/gemini/
- Public AI Tinkerers demos and talks: 15
- Result page: 1 of 1

## Recent Public Talks and Demos

### [From Meow to Mastery: Building MathCat’s Adaptive Learning Platform](https://hong-kong.aitinkerers.org/talks/rsvp_nk7cD0uzBpI)

MathCat.fun is an adaptive, bilingual math and science practice platform for Hong Kong primary school children, aligned to the EDB curriculum and playable as a Progressive Web App. For this demo, I’ll walk through the full stack, not just the game UI. Live, I’ll briefly show the working app (different play modes, leaderboard, rewards shop, etc); but will focus more on the backend: (1) the question-generation and validation workflow – the balance between deterministic and probabilistic models to achieve the best outcome and my learnings; (2) talk about the AI critic pipeline (Gemini + GPT-5) that scores each question template on Cantonese naturalness, HK school terminology, hint quality, and age fit (3) the adaptive progression architecture—mastery gates, recent-accuracy windows, skill unlocks, etc; (4) the overall architecture.

- Event context: AI Tinkerers Hong Kong: August Meetup with OAX Foundation and GMAsia — 2026-08-31 — Hong Kong
- Public talk page: https://hong-kong.aitinkerers.org/talks/rsvp_nk7cD0uzBpI

### [EvoFit: Building a Cross-Cultural AI Fitness Coach That Bridges Eastern Wellness and Western Exercise Science](https://hong-kong.aitinkerers.org/talks/rsvp_dvIFWxBl4TM)

EvoFit is an AI-powered fitness system that acts as a personalized coach — combining Eastern wellness traditions like Tai Chi with Western exercise science, all within a single app. Demo:For the demo: we're showcasing three live modules. The Movement Module lets users upload a workout video (e.g. squats) and receive AI-generated form feedback and scoring. The Real-Time Coaching Module uses your device camera for live Tai Chi practice — the system tracks your pose frame-by-frame, coaches you through movements, and scores your form in real time. The Food Module lets you photograph a meal to get instant calorie and nutrition analysis, or input your available ingredients and fitness goals to have AI generate a personalized recipe complete with instructions, macros, and who it suits best.

- Event context: AI Tinkerers Hong Kong at AWS: Agentic AI in Action (April) — 2026-04-29 — Hong Kong
- Public talk page: https://hong-kong.aitinkerers.org/talks/rsvp_dvIFWxBl4TM

### [Building NousyBooks - Orchestrating Low-Latency Multimodal Voice Agents with Gemini Live](https://atlanta.aitinkerers.org/talks/rsvp_La-nqq5lOEo)

I built NousyBooks, an AI-powered storytelling platform where children become the heroes of their own books. I built this project as part of Gemini Live Agent Hackathon Challenge. The core of the experience is "Nousy," a floating multimodal voice assistant that uses the Gemini Live API to brainstorm story themes, collect character details, and select art styles through natural, bidirectional conversation.

- Event context: AI Tinkerers Atlanta: Community Demos & Technical Deep Dives — 2026-04-21 — Atlanta
- Public talk page: https://atlanta.aitinkerers.org/talks/rsvp_La-nqq5lOEo

### [FuzzyClaw](https://cologne.aitinkerers.org/talks/rsvp_71A8QTW9ocg)

FuzzyClaw is an agent orchestration platform that turns markdown briefings into coordinated multi-agent work. A coordinator agent interprets the briefing, dispatches specialist agents running in isolated Docker containers, and routes their outputs into PostgreSQL, where they can be reviewed through a Django dashboard. Agents, the coordinator, and the human operator communicate in real time via a Redis-backed message board. The project is written in Python and draws inspiration from Steinberg’s OpenClaw and Qwibit.ai’s nanoclaw, with a strong focus on visibility, control, and practical day-to-day use.

- Event context: AI Tinkerers Cologne 4: Live Technical Demos — 2026-04-16 — Cologne
- Public talk page: https://cologne.aitinkerers.org/talks/rsvp_71A8QTW9ocg

### ["Saiba se sua campanha de mkt vai funcionar — antes de gastar R$ 1 em mídia."](https://saopaulo.aitinkerers.org/talks/rsvp_J6TWETVWK-U)

Ágora é um simulador de sucesso de campanha de mkt, com multiagentes que analisam, validam e comparam campanhas para identificar o que realmente gera resultado, aumentando a conversão de vendas sem depender de achismos. Através de um motor multi-agentes que funciona como um conselho de especialistas de marketing: 1. Um Analista Sociocomportamental que classifica sua campanha na era certa do marketing (1.0 ao 4.0), identifica de qual geração é seu público alvo (Z, Millennials, X, Boomers) e mapeia os vieses cognitivos a serem explorados. 2. Um Engenheiro de Oferta que desconstrói a promessa da sua campanha usando a matemática do Valor Percebido (Resultado x Probabilidade / Tempo x Esforço), apontando gargalos de fricção ou credibilidade. 3. Um Cientista de Dados de Performance que audita seus KPIs, punindo métricas de vaidade e definindo seu Timing Index para saber se "agora" é o momento certo para ir ao mercado. A Ágora oferece velocidade (análise em minutos, não semanas), previsibilidade (reduzindo em até 76% o custo de experimentação) e segurança para suas decisões, tudo isso com um custo até 90% menor que a pesquisa tradicional.

- Event context: AI Tinkerers SP e Banco BMG — 2026-03-26 — São Paulo
- Public talk page: https://saopaulo.aitinkerers.org/talks/rsvp_J6TWETVWK-U

### [AI-to-USD: An LLM Agent That Self-Corrects Industrial Scene Generation for Isaac Sim](https://nyc.aitinkerers.org/talks/rsvp_Y4sGAt1oy-A)

I built an end-to-end pipeline that turns natural language descriptions into validated USD industrial scenes and synthetic training data for computer vision. The pipeline has two stages: Stage 1: Spec-to-Sim Copilot: You describe a battery module assembly task ("assemble a 2x3 grid of LG E63 cells with a UR10e robot"). An LLM (Gemini 2.0 Flash with structured output) generates a Pydantic-validated ModuleTask JSON. Then, 5 industrial safety rules are checked against the spec: thermal spacing between cells, robot arm reachability, cell rotation alignment for busbar welding, module tray bounds, and cell count limits. If validation fails, the system feeds structured error reports back to the LLM for self-correction (up to 3 repair attempts). This is the interesting part: pure LLM generation fails silently for spatial tasks — cells overlap by 5-10mm, coordinates fall outside robot reach — so the validation loop catches what the LLM can't reason about. Stage 2: Industrial SDG Lab: The validated scene gets domain randomization across 4 axes (lighting intensity, material color shift, camera pose, object placement jitter) to produce variant USDA scenes with COCO-format annotations — ready for training object detection models. An Omniverse-style 3D viewport built with Three.js lets you interactively inspect the scene with PBR materials, bloom post-processing, and a property inspector panel. Live demo: I'll type a prompt, show the LLM generating a spec, trigger a validation failure, watch the AI self-correct, then generate randomized variants with the 3D viewport.

- Event context: March Demo Day, hosted by Flowglad — 2026-03-18 — New York City
- Public talk page: https://nyc.aitinkerers.org/talks/rsvp_Y4sGAt1oy-A

### [Building Document Consciousness: How I Taught Gemini to Think in 6 Dimensions](https://san-diego.aitinkerers.org/talks/rsvp_CuPjQckMXvM)

I'll demo Clasio, a document intelligence platform I built solo on a 100% Google Cloud stack, showing how I use Gemini 2.5 Flash (extensible to Gemini 3 in a few keystrokes) to extract what I call "6D Document Consciousness" - analyzing every uploaded document across What, Who, When, Where, Why, and How dimensions simultaneously. I received $25K in credits from Google Cloud for Startups for Clasio. The technical meat of the talk: - How I built an async AI queue that processes 25 documents in 75 seconds using 30 parallel Gemini workers on Cloud Run - The structured extraction prompt engineering that gets Gemini to reliably output 6D consciousness JSON (and what failed before it worked) - A 6-tier search waterfall that goes from exact consciousness match down to fuzzy vector similarity using pgvector on Cloud SQL - returning direct answers, not document lists - How I handle connection pool management when you have 30 concurrent Gemini API calls each taking 2-10 seconds (spoiler: release the DB connection before the API call, not after) - Live demo: upload a stack of immigration documents and watch the system classify, extract entities, detect deadlines, and answer natural language questions in real time This is a solo founder build - no team, no VC money, just a product person who learned TypeScript and shipped to production on GCP.

- Event context: AI Tinkerers San Diego: February Meetup at Google — 2026-02-27 — San Diego
- Public talk page: https://san-diego.aitinkerers.org/talks/rsvp_CuPjQckMXvM

### [LLM vs Classical Vision Models for Real-World Object Detection](https://dublin.aitinkerers.org/talks/rsvp_Gc378dJzNpk)

In this demo, I will show how I use an LLM for object detection in a Telegram Mini App, and compare this approach with classical vision models like YOLO. In the app, users need to take a photo of a physical gift before putting it into a public box. The photo is sent to the backend, where an LLM (Gemini 2.5 Flash-Lite) looks at the image and returns a simple result: object category, confidence, and a short description. The LLM does not make final decisions. Its output is checked by simple rules in the backend, which decide if the user can continue, if the gift should be blocked, or if an admin needs to review it. I will explain why I chose an LLM instead of a classical CV pipeline, how much it costs per request, how I handle failed model responses, and in which cases this approach works worse than models like YOLO.

- Event context: AI Tinkerers Dublin Meetup — Baseline, January 26, 2026 — 2026-01-26 — Dublin
- Public talk page: https://dublin.aitinkerers.org/talks/rsvp_Gc378dJzNpk

### [The Secret to Stunning UI: How AI Helped Me Design and Ship InkyCards in 7 Days](https://cologne.aitinkerers.org/talks/rsvp_EmnVHJCgoK0)

In just one week, I took InkyCard, a conversational language learning app from an idea in my head to a production-ready product. This session isn’t just about generating code; it’s about using AI to solve the "Developer Design Gap" and build a brand with a soul. I will demo the InkyCards workflow, focusing on: The Design Extraction Hack: How to feed UI inspiration (from sites like Dribbble) into LLMs to generate custom design systems, ensuring your app doesn’t look "AI-generated." Building the "Soul": Using Gemini Nano and Higgsfield to create a unique app mascot and custom iconography, overcoming creative blocks and building an emotional connection with users. Production Speed-running: A look at the "Plan-First" prompting strategy and context management that allowed me to build complex features—like real-time AI conversations (Firebase Vertex AI) and "Tap-to-Learn" flashcard generation—without losing code quality. Technical Deep-Dive: State Management & AI: How to use Claude and AntiGravity to scaffold architecture that stays clean as the project grows. The Fresh Convo Rule: My framework for managing LLM context to prevent "code rot" and quality degradation.

- Event context: AI Tinkerers Cologne #2: Let's Build. — 2026-01-21 — Cologne
- Public talk page: https://cologne.aitinkerers.org/talks/rsvp_EmnVHJCgoK0

### [Simplifying HSA Expense Tracking with Gemini](https://st-louis.aitinkerers.org/talks/rsvp_7-LaLcmekG0)

My solution takes receipts as input, and uses Gemini-2.5-Pro to extract the text from them and classify the expenses based on whether they are HSA qualified, then outputs JSON that can be easily stored in records meeting the requirements of an IRS audit. I will present the code I'm using, my model choices (and why I made them the way I did), and the prompts I engineered. I will put specific emphasis on how I use external category sets to teach Gemini how to recognize eligible products and the way my use of structured output and examples to improve accuracy and reliability.

- Event context: AI Tinkerers - St. Louis October Meetup — 2025-10-22 — St. Louis
- Public talk page: https://st-louis.aitinkerers.org/talks/rsvp_7-LaLcmekG0

### [Agentes de voz que modifican tu interface en tiempo real](https://medellin.aitinkerers.org/talks/rsvp_KLYwgATZGwk)

Cómo construí un sistema multi-agente de voz en tiempo real con el ADK de Google y Gemini Live API, aplicado a un caso real. Presento la arquitectura y el código de un asistente conversacional que entrevista al usuario por voz, invoca functions para buscar, analizar y comparar “traffickers” (media buyers) y recomienda el mejor perfil; todo full-duplex con audio de ida y vuelta, barge-in (interrupciones) y latencias bajas. La demo usa Gemini Live API para streaming bidireccional de audio y sesión persistente, y function calling para enlazar intenciones del modelo con acciones de UI (mostrar tarjetas, comparar perfiles, iniciar contacto). El front está hecho con Lit + Web Audio API + TypeScript, y el runtime con el Google GenAI SDK. Veremos: diseño multi-agente, tooling (declaración de funciones), manejo de sesión en vivo, codecs y sample rate (16 kHz in / 24 kHz out), y patrones para conectar tool calls con componentes visuales.

- Event context: AI Tinkerers Medellín #15 - Telescoped - 21 de Agosto, 2025 — 2025-08-21 — Medellín
- Public talk page: https://medellin.aitinkerers.org/talks/rsvp_KLYwgATZGwk

### ["Sports Videos Understanding at Scale" OR "Video Runner Bib OCR detection"](https://montreal.aitinkerers.org/talks/rsvp_IccvgKkVEhA)

there are two things I worked on in Q1 and 2024Q4 that might be interesting: 1. Using VLM to create a Visual Q&A using RAG and also using it for "guided" highlight detection 2. I also tried using the large context window offered in Gemini 2.0 Flash Thinking to read marathon runners' bib

- Event context: AI Tinkerers Montreal - May 2025 Meetup at LightSpeed — 2025-05-07 — Montreal
- Public talk page: https://montreal.aitinkerers.org/talks/rsvp_IccvgKkVEhA

### [Vibe coded Point of Sale iPad App](https://singapore.aitinkerers.org/talks/rsvp_HGIyM6AvdFs)

Used vibe coding to create an iPad App, no Swift experience no iPad OS experience. Its an MVP version that actually works for simple uses like selling items in a booth or a pop out.

- Event context: AI Tinkerers Singapore: 6th Meetup - April 25th, 2025 — 2025-04-25 — Singapore
- Public talk page: https://singapore.aitinkerers.org/talks/rsvp_HGIyM6AvdFs

### [Building educational content with AI workflows](https://toronto.aitinkerers.org/talks/rsvp_gNhQXjEyDJg)

We are working on AI assisted workflows to enable human content creators to be able to rapidly build educational courses that teach software development principles through a mobile application focused on optimizing for retention through learning science principles like spaced repetition (think Duolingo for software development). I will live demo as much of a full course creation workflow as I can fit in 6 minutes highlighting the various stages of our workflow, how the human can interact with the output and provide feedback at various stages, how we assess quality and validate correctness of the material. I will show how we organized the code for our workflows, which are all accessible currently as CLI tools, but we have future plans to productize the workflows for course creators.

- Event context: AI Tinkerers Toronto - April 2025 Meetup at Shopify — 2025-04-24 — Toronto
- Public talk page: https://toronto.aitinkerers.org/talks/rsvp_gNhQXjEyDJg

### [print("Welcome to AI Tinkerers Hong Kong!")](https://hong-kong.aitinkerers.org/talks/rsvp_ayqgPCAK2ac)

In this session, the Lead Organizer and Founder of AI Tinkerers Hong Kong will share insights from his experience attending an AI Tinkerers meetup in Toronto and share his motivation for bringing this initiative to Hong Kong. He will discuss the details of the initiative and outline how each meetup is hosted. This sharing would be done with an example python script, and we will use Google AI studio to view the presentation slides to explain the outline using Gemini 2 preview.

- Event context: AI Tinkerers - Hong Kong Meetup (December) - Inauguration — 2024-12-19 — Hong Kong
- Public talk page: https://hong-kong.aitinkerers.org/talks/rsvp_ayqgPCAK2ac

## Related Technologies

- [Flash](https://aitinkerers.org/technologies/flash) ([Markdown](https://aitinkerers.org/technologies/flash.md)) — 14 public demos
- [Claude Code](https://aitinkerers.org/technologies/claude-code) ([Markdown](https://aitinkerers.org/technologies/claude-code.md)) — 216 public demos
- [Gemini](https://aitinkerers.org/technologies/gemini) ([Markdown](https://aitinkerers.org/technologies/gemini.md)) — 188 public demos
- [GPT-5](https://aitinkerers.org/technologies/gpt-5) ([Markdown](https://aitinkerers.org/technologies/gpt-5.md)) — 30 public demos
- [PostgreSQL](https://aitinkerers.org/technologies/postgresql) ([Markdown](https://aitinkerers.org/technologies/postgresql.md)) — 106 public demos
- [TypeScript](https://aitinkerers.org/technologies/typescript) ([Markdown](https://aitinkerers.org/technologies/typescript.md)) — 205 public demos
- [FastAPI](https://aitinkerers.org/technologies/fastapi) ([Markdown](https://aitinkerers.org/technologies/fastapi.md)) — 181 public demos
- [LangChain](https://aitinkerers.org/technologies/langchain) ([Markdown](https://aitinkerers.org/technologies/langchain.md)) — 445 public demos
- [Python](https://aitinkerers.org/technologies/python) ([Markdown](https://aitinkerers.org/technologies/python.md)) — 662 public demos
- [Supabase](https://aitinkerers.org/technologies/supabase) ([Markdown](https://aitinkerers.org/technologies/supabase.md)) — 90 public demos
- [Anthropic](https://aitinkerers.org/technologies/anthropic) ([Markdown](https://aitinkerers.org/technologies/anthropic.md)) — 36 public demos
- [Artificial Intelligence](https://aitinkerers.org/technologies/artificial-intelligence) ([Markdown](https://aitinkerers.org/technologies/artificial-intelligence.md)) — 5 public demos
- [Backend](https://aitinkerers.org/technologies/backend) ([Markdown](https://aitinkerers.org/technologies/backend.md)) — 7 public demos
- [BERT](https://aitinkerers.org/technologies/bert) ([Markdown](https://aitinkerers.org/technologies/bert.md)) — 179 public demos
- [Claude](https://aitinkerers.org/technologies/claude) ([Markdown](https://aitinkerers.org/technologies/claude.md)) — 174 public demos
- [Claude Opus](https://aitinkerers.org/technologies/claude-opus) ([Markdown](https://aitinkerers.org/technologies/claude-opus.md)) — 18 public demos
- [Cloud Run Gen2](https://aitinkerers.org/technologies/cloud-run-gen2) ([Markdown](https://aitinkerers.org/technologies/cloud-run-gen2.md)) — 1 public demo
- [Cloud SQL](https://aitinkerers.org/technologies/cloud-sql) ([Markdown](https://aitinkerers.org/technologies/cloud-sql.md)) — 3 public demos