Production AI and security-critical systems, built by a senior engineer.

A production system for ingesting, indexing, and querying large-scale unstructured data — audio, text, transcripts — via natural-language LLM queries over the corpus. Architected and built from the ground up, deployed on-premise for a security-focused enterprise client with strict data-sovereignty requirements.

Designed and implemented the full pipeline end to end: ingestion, transcription, embedding, vector indexing, query orchestration, and role-based access. Inference runs entirely on local hardware via Ollama; analytics and event aggregations are split across Apache Druid and ClickHouse, with a vector store for similarity search over the embedded corpus. The system is proprietary and not publicly demonstrable.

A production booking platform for a private medical practice, delivered end to end in four days. Patients schedule and reschedule through natural conversation in a Telegram bot, backed by Gemini for natural-language understanding; staff review and override via a full admin panel.

The build is complete: Telegram bot interface, Gemini-powered scheduling logic, full admin dashboard, complete patient records management. The four-day timeline is the part the client talks about; the part I care about is that it shipped to production with everything a real medical practice actually needs around it.

Founded the company; architected and shipped an open-source, end-to-end encrypted photo backup application now serving 100,000+ users. Featured by Ars Technica.

Designed the full cryptographic protocol from first principles using libsodium — authenticated symmetric encryption, public-key cryptography, and key derivation — so that the server stores ciphertext and never holds material it can use to read user data. Cross-platform: iOS, Android, server, and infrastructure all built and operated by the founding team.