AI/ML Engineer — I take AI-powered applications from idea to production. My focus is integrating LLMs into systems that are robust, private, and maintainable: retrieval-augmented generation, on-prem inference, and the unglamorous engineering that makes model output reliable enough to ship.
I come from a backend background (C# / .NET, microservices, enterprise integrations) and I bring that discipline into AI work — schema validation, error isolation, observability, and tests.
I'm doing an MSc in Computer Engineering (2026–).
- On-premise RAG platform — question answering over company documents, backed by PostgreSQL + Qdrant vector search with local Gemma inference via Ollama. No data leaves the network.
- ollama-middleware — an OpenAI-compatible API gateway for locally hosted models:
/v1/chat/completions,/v1/embeddings, key auth, streaming, and model aliasing so the real model identity never leaks to clients. - Structured-output pipelines — turning free text into deterministic, schema-conformant JSON with validation, a normalization layer, and retry. Multi-tenant isolation by design.
- onmic.io — self-hosted, always-on WebRTC mesh voice rooms with screen sharing.
docker compose upbrings up the app, coturn, and automatic TLS.
Two of my pull requests are merged into Microsoft repositories:
- dotnet/maui — BlazorWebView URL parsing fix, shipped in the .NET 10 release
- microsoft/windows-drivers-rs —
cargo wdk newnow respects default VCS behavior
I founded @BuzzSpire, an open-source organization where I maintain Nexar — a modern, ergonomic HTTP client library for .NET with a fluent builder, interceptors, retries, and 95% test coverage.
- Languages — Python · C# · TypeScript · Rust · C/C++
- AI/ML — Ollama · Gemma · Qdrant · RAG · embeddings · prompt & pipeline design
- Backend — .NET · ASP.NET Core · Node.js · PostgreSQL · Docker
- Frontend — Next.js · React · Tailwind CSS