I am an AI engineer interested in the whole system behind a model: how it is trained, served, evaluated, grounded, and made useful in the real world.
At aXite Security Tools, I help build privacy-first AI for critical infrastructure. My work spans local model serving, agentic RAG, document intelligence, speech, evaluation, and the infrastructure that keeps every inference inside the customer's environment.
| Build Transformers, training pipelines, data curation, instruction tuning |
Serve vLLM, on-premise inference, OpenAI-compatible APIs, GPU systems |
Evaluate Reproducible benchmarks, telemetry, retrieval quality, uncertainty |
Local AI for organizations where data cannot leave the building. I contributed to a production platform that combines private chat, a custom multi-agent RAG system, meeting transcription, document OCR, coding assistance, and continuous model evaluation. The stack runs on owned infrastructure with no external inference dependency.
on-premise AI agentic RAG vLLM FastAPI Qdrant PostgreSQL Redis Docker
A living benchmark for the models we actually serve. Twelve frozen tasks preserve original artifacts, side-by-side outcomes, Fable scores, and measured vLLM telemetry in a public, reproducible evaluation surface.
LLM evaluation vLLM telemetry Next.js reproducibility
A 125M-parameter language model trained end to end. I curated and tokenized an 18B-token, 120GB corpus, built memory-mapped data loading, and trained across 8x A100 GPUs in roughly eight hours. I then instruction-tuned the model and evaluated it against the original GPT-2 with EleutherAI's harness.
Training write-up / Pretrained model / Instruction-tuned model / Implementation
Turning 3D player tracking into coaching feedback. At the Ajax Hackathon 2026, our team analyzed frame-level skeleton data from a live Eredivisie match, isolated 23 shots, and built a kinematics engine that scores how efficiently momentum travels through a player's shooting chain.
sports analytics 3D kinematics biomechanics data visualization
- World Cup Agents - seven frontier models forecast the 2026 tournament and manage virtual $1M bankrolls.
- Jobfinder - discovers roles, ranks fit, sends alerts, tracks applications, and drafts tailored cover letters.
- Agent Trading - turns public insider and congressional disclosures into transparent paper-portfolio decisions.
- DutchInquire - retrieval and question answering over Dutch-language documents.
Uncertainty Estimation for Super-Resolution Using ESRGAN
Maniraj Sai Adapa, Marco Zullich, and Matias Valdenegro-Toro. VISAPP 2025.
We combined ESRGAN with Monte Carlo Dropout and deep ensembles to produce calibrated, per-pixel uncertainty maps without sacrificing super-resolution performance.
Based in the Netherlands. Building private AI systems from research to production.