AI/ML Engineer · MSc Computer Engineering — I build AI systems that run on your own hardware and produce output reliable enough to ship.
Backend engineer by background (C# / .NET), now working on Retrieval-Augmented Generation, LLM integration, on-premise inference and vector search. Most of what I build answers the same question: how do you get useful, verifiable output out of a language model without a single byte leaving the network?
Currently doing an MSc in Computer Engineering (2026–).
- RAG & vector search — document chunking, embedding stores (Qdrant, PostgreSQL), semantic retrieval pipelines
- On-prem inference — Ollama, LM Studio, Gemma; sensitive data stays inside the network
- Structured-output reliability — strict JSON schemas, validation, a normalization layer that recovers malformed model output, and retries
- LLM serving infrastructure — OpenAI-compatible gateways, model aliasing, multi-key auth, streaming, embeddings endpoints
| Contribution | What it was |
|---|---|
| dotnet/maui | BlazorWebView URL parsing fix. Path.HasExtension() was run against the whole URL, so a query parameter like ?weight=62.5 looked like a file extension and blocked force-reload; the fix checks only the path component. Shipped in .NET 10. |
| microsoft/windows-drivers-rs | cargo wdk new now respects default VCS behavior — removed a hardcoded --vcs none. Merged Nov 2025. |
| UI5/webcomponents | Theming fix for read-only TextArea focus and border. |
I also run BuzzSpire — an open-source organization, not a commercial company (I own the trademark; the org itself is a community project). Its main library is Nexar, an HTTP client for .NET: fluent request builder, typed responses, interceptors, exponential-backoff retries and auth helpers. 76+ tests, 100% pass, 95% coverage, MIT.
| Project | What it is | Stack |
|---|---|---|
| Internal LLM Platform | Fully on-premise RAG assistant answering questions over company documents. Qdrant + PostgreSQL embedding store, chunking and semantic retrieval, local Gemma inference, API key management, one-command Docker Compose deploy. | .NET 9 · Next.js 15 · Qdrant · Ollama |
| ollama-middleware | OpenAI-compatible gateway in front of local models — /v1/chat/completions, /v1/completions, /v1/embeddings, /v1/models, Bearer auth, streaming, sampling passthrough. Model aliasing keeps the underlying model identity out of every response. |
Node.js · TypeScript |
| Teams AI reporting pipeline | Meeting transcripts → deterministic, schema-conformant JSON reports (summary, decisions, action items, risks). Multi-tenant isolation, on-prem inference. | Next.js · Ollama · SQLite |
| On-prem assistant for OpenText SMAX | Manifest V3 Chrome extension embedding a local LLM inside an ITSM interface. Zero-click ticket context capture, prompt-injection hardening across three isolated JS execution contexts. | TypeScript · React · CRXJS |
| neuralnetwork3d | Watch neural networks train in live 3D in the browser — zero dependencies. | Python · HTML |
| onmic.io | Self-hosted, always-on WebRTC voice + screen share over a peer-to-peer mesh — no media server. coturn TURN for clients behind NAT, Caddy for automatic TLS, single docker compose up. |
TypeScript · React · Fastify |
| csharp-ile-programlama | The 232 code listings from a C# / .NET 10 book I wrote (41 chapters, not yet published). Every listing is syntax-checked with Roslyn in the build pipeline. | C# |
| OCRmyPDF-Qt-GUI-Client | Multi-threaded Qt GUI for OCRmyPDF, 100+ languages. | Python · PyQt5 |
- AI/ML — RAG, LLMs, embeddings, vector search, semantic retrieval, structured-output validation, on-prem inference, Ollama, Gemma, Qdrant, PyTorch
- Languages — Python, C#, TypeScript, JavaScript, Rust, C/C++, SQL
- Backend — .NET 8/9, ASP.NET Core, EF Core, Node.js, microservices, REST, YARP API gateway, JWT auth, Clean Architecture, DDD
- Data & infra — PostgreSQL, SQL Server, Qdrant, Redis, Elasticsearch, Docker, Docker Compose, GitHub Actions, Cloudflare Tunnel
- Frontend — Next.js, React, TypeScript, Tailwind CSS
- Enterprise — Siemens Teamcenter PLM (ITK, SOA, Active Workspace), SAP T4S, OpenText SMAX, Active Directory / LDAP, Microsoft Graph
Backend and enterprise integration, on-site in İstanbul: secure integration services between Siemens Teamcenter PLM, SAP and ITSM systems — Teamcenter server extensions and custom handlers in C/C++ (ITK), SOA and REST services in C# / .NET, a YARP-based API gateway with JWT auth, AD/LDAP and Microsoft Graph user sync, PostgreSQL and SQL Server, Docker and CI/CD. The AI work above is all personal.
- Website — mertturkoglu.com
- LinkedIn — in/18w