Projects & experiments

AI motion graphics for Hebrew short-form video

Turns a Hebrew talking-head video into a finished short with explanatory motion graphics timed to the exact word being said. Whisper speech-to-text (an ivrit.ai Hebrew model) gives per-word timings, Claude reads the transcript to decide what to draw and writes each animation as code, and YuNet face detection keeps graphics clear of the speaker. Runs on CPU with no GPU, with a web editor that composites the graphics live over the video.

Hebrew mortgage-advisory platform for a live finance business

Right-to-left Hebrew platform where borrowers answer a branching questionnaire and get a scored assessment, a document portal and a paid mortgage-mix recommendation. A GPT-5.1 agent with structured outputs (strict JSON schema) splits the variable-rate tracks inside deterministic guardrails, with a rule-based fallback and versioned prompts tested in an admin simulator. FastAPI and Supabase Postgres, with payments, CRM sync and ~800 tests.

AI tax platform, VLM extraction + RAG chatbot

The AI tax platform where I led full-stack. A VLM engine converts multilingual tax forms into structured U.S. tax data, LLM-powered assistants automate support and tax workflows for customers and internal teams using RAG over vector search (LangChain + FAISS), and a tax optimization engine analyzes financial data to find savings. Also improved platform performance by 40%.

Survival-chase browser game with a bot that plays it

A cursor-controlled 2D survival game on HTML5 Canvas: you're a mouse dodging a cat driven by predictive game AI that powers up with special attacks, across three modes with online leaderboards. Every visual is drawn and every sound synthesized with the Web Audio API. The server replays each run's inputs to verify the score, and a search-based bot simulates ~100 candidate moves per frame to play to any target score.

Autonomous AI news publisher for LinkedIn

Every morning it pulls headlines from 10 RSS feeds and sends them to Telegram. For each one I approve, Claude writes the LinkedIn post and a local Z-Image Turbo diffusion model (run through ComfyUI, with OpenAI image generation as fallback) illustrates it. A second tap publishes it to the company page through the LinkedIn API.

Stops double-redeemed coupons at Pinat HaOfe bakery

Sellers photograph the customer's coupon screen as they already did, and the photo itself becomes the redemption. Google Vision OCR returns every word with its coordinates, a label-proximity algorithm picks the coupon code out of up to 16 competing numbers on the receipt, and a burn list answers with a green or red tone in under a second. Offline queue, manager dashboard with duplicate diagnostics, CSV export.

Bilingual booking site for a private tutor

A bilingual (Hebrew / English) site where a private tutor's students book Hebrew-language and math lessons. Built with React + Vite + Tailwind and deployed end to end on a self-managed cloud instance with nginx, TLS and self-hosted, cookie-free analytics.

Bilingual parimutuel betting pool (play-money)

A Hebrew-RTL + English betting pool where friends spend free credits predicting a baby's gender, birth date, weight and more. Each market is a parimutuel pot split among the closest guesses, with all credit-minting locked down in Postgres Row-Level Security and SECURITY DEFINER functions behind an append-only ledger.

Generative AI that builds complete branded websites

At OpenLoop Health (work, private): a generative AI platform for GLP-1 and other medication programs that uses LLMs and image models to create branded websites, copy, logos and visual assets. A VLM engine analyzes visual inputs into structured context for generation, and an AI assistant with tool calling guides users through site creation and refines the generated content.

Realtime Hebrew voice AI agent

A phone and browser agent that holds live Hebrew conversations, qualifies leads and books meetings. Audio streams over WebSockets through ivrit.ai Whisper speech-to-text, OpenAI Realtime and Cartesia text-to-speech, while a LangGraph graph of Pydantic AI agents (intent, profile, pitch, scheduling) steers the call and a guardrail agent rewrites every reply before it is spoken. Langfuse traces each session.

Pluribus-style 6-max poker AI

Team entry for a bot-vs-bot No-Limit Hold'em competition. A Deep CFR self-play blueprint (PyTorch neural networks) feeds runtime lookup, an exact real-time CFR search on the river and a per-opponent exploitation layer, routing every decision to the layer that plays it best. A Gemini vision reader powers a coach that reads live tables from screenshots. 162 tests across 24 files.

LLM market-signal alerts on Telegram

Watches a public figure's Truth Social posts with a headless browser and runs each new one through GPT-4o for a market-impact analysis: a bullish/bearish rating, a -10 to +10 score and the affected stocks and sectors, returned as structured JSON and pushed to a Telegram channel in near real time.

Browser eye-tracking heatmaps

Proof of concept that tracks where users look through the webcam: a MediaPipe FaceMesh model with iris tracking locates the pupils in the browser, a 9-point calibration maps them to the screen, and a Dockerized Flask backend renders a live attention heatmap.

AI Buddy · TLDR Buddy · ChatGPT timestamps

Chrome extensions for working with AI in the browser. AI Buddy gives AI summaries of any highlighted text and rewrites form input in a chosen tone with GPT-4o, called through a Cloudflare Worker proxy so no API key ships in the extension; TLDR Buddy was its first version. A third extension adds persistent timestamps to every ChatGPT message.

Fair group game-picker with carry-over voting

Helps a group fairly choose what to play next: votes for games that don't win carry over to future rounds, so minority preferences eventually win. Google sign-in rooms with Steam game search, live updates over Server-Sent Events, and one vote per round enforced by a database index.

Coding-interview prep, read aloud

A mobile app for studying coding-interview problems hands-free: 151 NeetCode 150 / Blind 75 questions across 18 topics, tagged by difficulty and read aloud with adjustable-speed text-to-speech. Built with Expo / React Native.