Martin Kaiser

AI/ML Engineer · PhD Physics · Berlin

I build LLM applications, agents and document pipelines.
Featured: Büchli — 0.930 macro-F1 on 976 held-out claims from 14 unseen vote dates

PhD in Physics · 3+ years shipping LLM systems · 7 hackathon wins · 12 podiums (2025–2026)
mloda open-source plugin grant winner · 3,100+ contributions in the last 12 months · 117 public repos

🎯 Looking for: AI/ML engineering roles · agent & LLM product teams · contract work — Berlin or remote

role:        AI/ML Engineer
location:    Berlin, Germany 🇩🇪
focus:       LLMs · Agents · Automation · Document Processing · Edge AI
status:      Open to roles, projects, collaborations & hackathons
background:  PhD Physics → 3+ years shipping LLM systems

_{Also in 2026: 4th place — Superchat × Needle AI Builder Event (Salesbot) · Top-7 finalist — Cognee AI-Memory Hackathon}

_{A slice of 160+ repos (117 public) — grouped to show the range.}

🧩 Agents & LLM apps

• reaction-dynamics — 🥇 team-behavior analytics from live Slack reaction timing; KS-test classifier + Cognee/Qdrant knowledge graph
• Soteria — 🥉 federated multi-party incident response (Flower ServerApp + ClientApp + AgentApp); parties answer with traffic lights, never raw values; tiered measures, two-key human approval, chained receipts
• AngebotsBot — AI deal scanner: kaufda.de + Qwen VL 32B + Supabase + LangGraph + Chainlit + Telegram
• fraudmind — deterministic fraud-hunting over GDPdU audit dossiers; 22 rule families + Cognee knowledge graph, "no number without a source"
• carlover — multi-agent automotive assistant (LangGraph + ADAC data + Featherless AI)
• multi-agent-research — research automation with LangGraph + Claude + SerpAPI

🔬 Evaluation & safety research

• hack-apertus-booklet-nli — multilingual NLI with evidence: 16 retrieved passages beat the whole booklet on accuracy and cost; bootstrapped CIs over vote dates
• do_models_have_minds — content control for preference-coherence metrics (Apart Digital Minds Sprint)
• secret-localities-strategies — activation-forensics pipeline for model organisms with hidden loyalties (Apart Secret Loyalties hackathon)
• sayproof — "no source, no verdict": verbatim-quote guard on LLM fact-checking

⚡ Edge & self-hosted AI

• jetson-headless-inference — Jetson Orin 8GB headless AI API with model switching & power management
• comm-buddy (private) — local-first meeting analysis: whisperX + pyannote + Ollama across Mac + Jetson over Tailscale
• mac-ram-manager · cleandrive — zero-dependency macOS system tools

Claude Code is my daily driver. I pair it with memory frameworks like claude-mem and agent harnesses like ECC and superpowers to reuse context and cut redundant re-reads, and I measure the result instead of guessing.

I instrument my own sessions with claude-code-langfuse-tracing; these are the numbers from that tracing for 30 days up to September 2026:

Things I've shipped for the Claude Code ecosystem:

• 🦊 kitsune-mcp — shape-shifting MCP hub: shapeshift() into 10,000+ MCP servers at runtime, one entry point, no restarts
• 🔭 claude-code-langfuse-tracing — zero-intrusion observability; a Stop hook replays transcripts into Langfuse as backdated, session-grouped traces with token/cost
• 📡 claude-A2A-Comm — A2A-protocol messaging between Claude Code sessions, local or over a Pi/VPS hub
• 🧰 marty-skills · kogitsune — personal skills, commands & harness for Claude Code

Core

AI & GenAI

Data & Infra

Edge

🎤 VP Public Relations · Toastmasters Club Berliner Rhetorikmeister
♟️ Teams Organizer · Chessclub Weisse Dame e.V. (200 members)

_{Hover every day and drag the inputs to see how the rank responds: interactive version}

_{💬 Berlin-based · open to AI/ML engineering roles, contract work and hackathon teams — martinkaiser.bln@gmail.com}