I build LLM agents and evidence-bound document pipelines — and measure whether they work.
| 🏆 7 wins · 13 podiums | 🦊 Awesome MCP Servers | 🔬 Top 10% | 💰 Grant winner |
|---|---|---|---|
| hackathons 2025–2026 · 15 top finishes | kitsune-mcp listed · ~2.8k downloads/mo | Apart Research sprint, 230+ projects | mloda open-source plugin grant |
🎯 Open to: hackathon teams · open-source collabs on agents & LLM tooling — Berlin or remote
🦊 kitsune-mcp — hot-reload for MCP servers
Your agent edits an MCP tool, calls reload("dev") and tests the fix in the same session, with no client restart. It also mounts any of 130,000+ servers from 7 registries mid-session, and runs community packages inside a hardened Docker cage.
Merged into punkpeye/awesome-mcp-servers (97k★, Oct 2026) · v0.22, 270+ commits, CI + coverage · claude mcp add -s user kitsune -- uvx kitsune-mcp
| Tool | What it does |
|---|---|
free-llm-proxy-router |
One Go binary turns every free LLM tier (Groq, Gemini, OpenRouter, Cerebras, GitHub Models…) into a single OpenAI-compatible endpoint with automatic fallback |
claude-code-langfuse-tracing |
A Stop hook replays Claude Code transcripts into Langfuse as backdated, session-grouped traces with token and cost; fail-open, secret in the macOS Keychain |
| Astrolab · ▶ Live | Your GitHub stars as a decision tool: 2,400+ starred repos in a 3D knowledge graph (39 communities), health and lifecycle scores, and 29 landscape reports that answer "which tool should I use for X?". Built offline, no LLM in the hot path, refreshed weekly |
| Project | Why it matters | ||
|---|---|---|---|
| 🥈 | Friends with Benefits 2nd · Stripe × Supabase Hack Berlin · Oct 2026 |
One sentence from a shop owner becomes a booking page with an ambassador crew. Rewards fire only on a real Stripe payment, so fake sign-ups farm nothing | ▶ Live |
| 🥇 | reaction-dynamics 1st · Cognee × Qdrant Hack Night · solo |
Team-behavior analytics from Slack reaction timing: KS-test classifier + Cognee/Qdrant knowledge graph | ▶ Live · graph |
| 🇨🇭 | Büchli Hack Apertus 2026 · Track 2A · Oct 2026 |
Checks campaign claims against the Swiss voting booklet in DE/FR/IT with Apertus 70B. 0.930 macro-F1 on 976 claims from 14 unseen vote dates, at 1/9 the tokens of whole-booklet prompting | |
| 🏅 | GLiNER2 action classifier (private) Top 10% · Apart AI Incident Response Sprint · Sept 2026 |
A fine-tuned 307M encoder flags agent action commitments in 18 ms vs 520 ms for a hosted Qwen baseline, catching 14 of 15 actions vs 10 — a high-recall tripwire for agent monitoring | |
| 🎤 | Sayproof Nebius × NVIDIA Global AI Hackathon · Sept 2026 |
Fact-check a talk before the audience does: every verdict quotes its source verbatim, or says "can't verify". Nemotron 120B + Tavily | ▶ Live |
More live: fraudmind — fraud hunting over GDPdU audit dossiers, 22 rule families · Soteria on Monad — incident receipts on chain · LLM pricing audit
7 wins · 13 podiums · 15 top finishes · 2025–2026
| Result | Event | Project / Track |
|---|---|---|
| 🥇 1st place | Cognee × Qdrant Hack Night · Berlin | reaction-dynamics — solo entry |
| 🥇 1st place | Agentic FinAI Hackathon · Berlin | Dify Open Track |
| 🏆 Track winner | AI Agents Hackathon · Berlin | Blockchain for Good Alliance Bonus Track |
| 🥈 2nd place | Stripe × Supabase Hack: AI Payments for the Real Economy · Berlin | Friends with Benefits |
| 🥈 2nd place | AI Builders Open Claw Hackathon | Open Claw |
| 🥉 3rd place | Collaborative Agent Hackathon · Flower Labs, Berlin | Soteria — privacy-preserving multi-party incident response |
| 🥉 3rd place | Needle Hackathon | Main Track |
| 🏅 Runner-up · 4th place | Superchat × Needle AI Builder Event | Salesbot |
| 🏅 Top-7 finalist | Cognee AI-Memory Hackathon | Finals |
| Result | Programme | Project |
|---|---|---|
| 🏅 Top 10% of 230+ | Apart AI Incident Response Sprint · Containment · Sept 2026 | GLiNER2 as a fast action classifier for agent monitoring (private) |
| 💰 Grant winner | mloda monthly plugin grant · Sept 2026 | mloda-gdpdu-provenance (private) — a total that names its rows, or no total |
| 📝 Submitted | Apart AI Collusion Sprint · Track 2 · Oct 2026 | auditing-llm-pricing-agents — legal price rules mimic the punish-and-recover pattern collusion audits look for |
| 📝 Submitted | Apart Digital Minds Sprint · Aug 2026 | Coherence Without Content — a missing content control for preference-coherence metrics |
| 📝 Submitted | Apart Secret Loyalties Hackathon · Berlin, July 2026 | secret-localities-strategies — activation forensics for hidden loyalties |
2025 — 4 wins · 6 podiums (click to expand)
| Result | Event | Project / Track |
|---|---|---|
| 🥇 1st place | ElevenLabs Hackathon | n8n Category |
| 🥇 1st place | NEAR AI Hackathon | Shade AI-Agent |
| 🥇 1st place | AI Hardcore Hackathon | LLM Prediction |
| 🥇 1st place | Masterschool Hackathon | SMS-Service |
| 🥈 2nd place | Distributed AI Hack Berlin | Open Track |
| 🥈 2nd place | IronHack DevOps Hackashow | DevOps |
A slice of 189 repos (128 public).
🧩 Agents & LLM apps
Soteria— 🥉 federated multi-party incident response on Flower; parties answer with traffic lights, never raw values; two-key human approval, chained receiptsfraudmind— deterministic fraud hunting over GDPdU audit dossiers; 22 rule families + Cognee graph, "no number without a source"AngebotsBot— AI deal scanner: kaufda.de + Qwen VL 32B + Supabase + LangGraph + Telegramcarlover·multi-agent-research— LangGraph multi-agent assistants
⚡ Edge & self-hosted AI
jetson-headless-inference— Jetson Orin 8GB headless AI API with model switching & power managementcomm-buddy(private) — local-first meeting analysis: whisperX + pyannote + Ollama across Mac + Jetson over Tailscalemac-ram-manager·cleandrive— zero-dependency macOS system tools
📊 ML & benchmarking · 🔁 Automation & infra (click to expand)
📊 ML & benchmarking
featherless-bench— systematic LLM benchmarking + DeepEval; Pareto frontier, Vision Arena, 19 text modelsllm-finetune-kit— beginner-friendly LLM fine-tuning: 3-line training, web UI, smart defaultsberlin25-eurosat— federated learning (Flower) on EuroSAT satellite imagery
🔁 Automation & infra
inbox-inferno-n8n— AI email agent with LLM-as-judge — scored 20/20aws-terraform-devops-infrastructure— AWS VPC infrastructure defined with Terraformkaiser-echo— bilingual voice agent with memory (React + Cloudflare Workers)
Claude Code is my daily driver. Each layer below does one job, and I measure the result instead of guessing:
| Layer | What it does in my sessions |
|---|---|
| superpowers | The process: brainstorm → written plan → test-first → verify with evidence before anything is called done |
| ECC | Code-quality rules for Python and TypeScript, switched on per session when I write or review code |
| claude-mem | Memory across sessions: what was tried, decided and shipped, so a new session starts where the last one stopped |
| 32 public skills of my own | Built with my skill-forge, mostly research methodology learned the hard way in eval sprints: systematic-study, seed-replicates-before-effects, detection-needs-a-denominator, numbers-from-primary-sources |
Numbers from my own Langfuse tracing, 30 days to 9 Nov 2026:
| Traces | Observations | Tokens (30d) | Model |
|---|---|---|---|
| 5.6K | 67K | ~4.0B | Opus 5.5 for 99.9 % of tokens |
- 🦊
kitsune-mcpand 🔭claude-code-langfuse-tracing— see Open source in use - 📡
claude-A2A-Comm— A2A-protocol messaging between Claude Code sessions, local or over a Pi/VPS hub - 🧰
kogitsune— my session launcher: starts each project with a lean kit of skills and plugins, so sessions spend fewer tokens
🎤 VP Public Relations · Toastmasters Club Berliner Rhetorikmeister
♟️ Teams Organizer · Chessclub Weisse Dame e.V. (200 members)
-3,800+_·_top_1%25-00ff88?style=for-the-badge&logo=github&labelColor=0d1117){kind=icon}
Hover every day and drag the inputs to see how the rank responds: interactive version
How the rank is computed (click to expand)
The rank uses the formula from github-readme-stats, ported unchanged. Each input is scored against a fixed median and weighted; the weighted share sets the percentile and the letter.
| Input | Median | Weight |
|---|---|---|
| Commits, last 12 months | 250 | 2 |
| Pull requests, all time | 50 | 3 |
| Issues, all time | 25 | 1 |
| PR reviews, last 12 months | 2 | 1 |
| Stars on owned repos | 50 | 4 |
| Followers | 10 | 1 |
Levels by percentile: S ≤ 1 · A+ ≤ 12.5 · A ≤ 25 · A- ≤ 37.5 · B+ ≤ 50 · B ≤ 62.5 · B- ≤ 75 · C+ ≤ 87.5 · C otherwise.
What the letter does not measure: its "percentile" compares inputs against fixed medians, not against real users, and stars carry a third of the weight, so activity alone tops out around A-. For a measured comparison, 400 random GitHub accounts were sampled (Nov 2026): none had as many contributions in the last 12 months as this profile, which puts it in the top 1 % at 95 % confidence.
