Architected and lead-developed the Swiss AI Hub, an enterprise AI platform now released as open-source under Apache 2.0. ~35 integrated containers — LLM gateway, vector search, multi-agent runtime, data pipelines, SSO, and full observability — deploy with one command, run air-gapped on a single GPU, and ship across 15 projects for 12 Swiss enterprise clients.
Built a Swiss-language PII corpus (212k chunks across de_CH, fr_CH, it_CH, rm, en) from a three-LLM consensus plus checksum-validated regex, then fine-tuned xlm-roberta-large on it. 0.910 strict-span F1 on the held-out test set against 0.44 for the next-best public detector.
Fine-tuned a 0.6B parameter LLM to restore flattened heading hierarchies in markdown extracted from PDFs. 80.6% per-heading accuracy, runs on CPU, published as a PyPI package and HuggingFace model.
Sends a swarm of Claude Code agents to analyze any codebase from every angle — architecture, git history, patterns, tooling, security — then synthesizes a production-grade configuration with CLAUDE.md files, skills, hooks, subagents, and MCP servers.
Built an MCP server that turns any AI assistant into a presentation tool. Ask it to create slides, pick a theme, and get a shareable link that works in any browser. 24 themes, iterative editing, 30-day hosted URLs.
Built an MCP server that gives AI assistants direct access to live Vue ecosystem documentation. Hybrid retrieval over 8 frameworks, structure-aware chunking, deterministic query pipeline, free hosted at mcp.vue-mcp.org.
.