My Projects

Decodes speech from ultrasound video of the tongue, letting people talk to a computer without making a sound. Built in one month: 15.6% WER on open-vocabulary silent speech from 50 hours of data, generalizing to unseen speakers. Published with Aleph Neuro. Blog > Demo >

Licenses 12PB of multimodal training data for AI: conversations, speech, video, images, and documents. The collection includes 261B+ messages across 50+ languages, natively entangled across modalities, and is used by frontier AI labs and Fortune 500 companies. Website >

Party game with 2.5M+ downloads, regularly ranking top-5 among games in Eastern Europe. Google Play > App Store >

Crypto trading system that dynamically rebalances across pairs and strategies to capture market inefficiencies in real time. Meta-simulators continuously tune market parameters. Running since 2018, with a 52x return to date. Private project.

One of the first PyTorch implementations of YOLOv4, extended with 20+ training and inference features. Developed with guidance from Alexey Bochkovskiy. GitHub >

Solana trading extension that executes trades faster than block finalization. Website >

Personal AI assistant built on Claude Code, with a scheduler, Telegram bot, dashboard, task queue, heartbeat, skills, and integrations across Gmail, Calendar, GitHub, Obsidian, and messaging platforms. Blog >

Personal data lake that syncs 19 sources into a unified FTS5 search index and provides the memory layer behind Klava. GitHub > Blog >

Native macOS meeting recorder with on-device transcription and AI notes. Built in four hours as a private, subscription-free alternative that works across languages. GitHub >

Generated ads featuring celebrities, with royalties paid per impression. App Store > Website >

Mobile app for discovering nearby museums and taking interactive audio tours. App Store > Website >

Computer vision system for detecting weeds among crops, including assisted plant annotation and a complete YOLOv4 training and inference pipeline.

Model that detects, classifies, and localizes animals in scouting-camera video with 96% accuracy, using GradCAM for interpretability and specialized video preprocessing.

Second place at AGI House Hackathon. The app surfaces agreements, disagreements, and contradictions across news sources. GitHub >

Hackathon prototype using license-plate and vehicle recognition to flag mismatched vehicles and route incident reports to city officials. DevPost > Reddit >