Mathematician by training (PhD in extreme value theory), full-stack AI engineer in practice. I like problems that need both a whiteboard and a deploy pipeline.

🔭 These days I'm co-founder & CDO at Nabu, where I've spent the last six years building the data and ML side of a document-intelligence platform for customs & trade — turning messy, unstructured trade documents into structured, calibrated data. In practice, a lot of slow manual processing collapses into a few minutes.

• Document AI & RAG — LangChain / LangGraph, Weaviate & pgvector, visual-rich document understanding, custom OCR pipelines
• LLM systems with some rigour — dynamic structured outputs, logprob-calibrated confidence scores, and MILP where it earns its keep
• The stack around it — FastAPI, PostgreSQL, React/Vue, AWS (EKS, SageMaker), Terraform, Kubernetes, and a soft spot for observability (OpenTelemetry, Datadog, SigNoz, Grafana)

• 📈 Quant & crypto tinkering — backtesting ideas, poking at market data, and the occasional Ethereum rabbit hole
• 🧪 ML / RL & generative AI for fun — reinforcement learning, self-hosted LLMs, Stable Diffusion
• 🏠 Self-hosting & homelab — Proxmox, ZFS, Grafana dashboards, and a healthy distrust of the cloud for personal stuff
• 🗃️ Data hoarding — web archiving, media library tooling, metadata wrangling, giving everything a tidy, well-tagged home. If it can be catalogued, I've probably tried.

• Beets-Plugin_VGMdb — VGMdb metadata for the beets music manager
• fast_deskew — a fast document deskew library, born from real OCR pain
• py-prisma2markdown — turn Prisma schemas into readable Markdown docs

⚡ Fun fact: all of this runs on three servers and ~200 TB at home — which I assure everyone is "for the homelab" and definitely not just hoarding.

💬 Always happy to talk document AI, applied ML with real math behind it, quant experiments, or homelab over-engineering.

📫 hozhenwai@gmail.com · LinkedIn · based in Strasbourg 🇫🇷