Selected work

Live and historical European airspace: ADS-B positions, OpenSky movements, schedules and weather stream through Kafka into a Bronze and Silver lake, a dbt warehouse on DuckDB and a read-only serving copy behind a public dashboard.

GDPR-native streaming and medallion lakehouse for European commerce: real-time fraud scoring, DuckDB and Turso serving, and a six-layer Article 17 erasure cascade with a verifiable audit trail.

A terminal UI and CLI uninstaller for Linux that goes further than package managers do on their own. It scans 17+ package managers at once (pacman/AUR, apt, dnf, Flatpak, Snap, AppImage, npm, pipx, cargo, and others), finds the leftover config, cache, and state directories that normal uninstalls leave behind, and removes them in one transaction behind a single sudo prompt.

Privacy-first, self-hosted writing assistant: a browser extension with a 156,000-word offline engine and more than 40 rule patterns, an AI router that only sees what local rules miss, and a Hono backend that deploys to five runtimes.

A multi-camera surveillance and behavioral monitoring platform that runs on ordinary hardware. It does real-time face recognition with anti-spoofing liveness checks, tracks occupancy and gate traffic, remembers where tracked assets were last seen, sends alerts, and includes an AI copilot that answers questions about the live database in plain English.

OpenCodeHub is a self-hosted, open-source Git platform meant as an alternative to GitHub and GitLab. Git hosting, pull requests, stacked PRs, merge queues, CI/CD, AI-assisted code review, issues, organizations, webhooks, REST and GraphQL APIs, and a CLI all live in one deployable system. It's built as a modular monolith with optional background workers and CI runners, so it runs on one machine and can still scale horizontally. Git works over HTTP Smart Protocol and SSH. Storage is pluggable (local or S3-compatible), several database backends are supported, CI runs in Docker containers, queues use Redis, and code review can talk to multiple AI providers.

A modular reasoning architecture that knows when to abstain: six stages with an explicit uncertainty channel, a deterministic DAG router, stateless experts and a verification anchor. GSM8K accuracy 47.2% to 55.6% while confidently wrong answers drop from 19.8% to 2.9%.