Berlin · ex-COO of Fyrfeed (acquired 2024) · founder of ai1 Ventures
I'm an operator who ships production AI systems — I direct them end-to-end with Claude Code rather than hand-writing them. The through-line across everything below: agents that must cite their sources, pass tests in CI, and refuse rather than guess when they can't back a claim. What that demonstrates is deployment judgment — grounding, eval gates, and refusal-by-construction — shown in code I directed rather than typed.
Right now: an autonomous research-to-schedule content pipeline shipping short-form video daily — ~22 videos/week across TikTok, Instagram, Facebook & YouTube, unattended behind human approval gates (research → script → render → schedule, built on Claude Code) — plus an installable MCP server for running AI transformation programs, and the repos below.
| ai-transformation-90 | A transformation copilot for the first 90 days of an AI program — playbook, costed business cases, working prototypes, and the method itself as an MCP server with evidence-gated scoring. Swap in your own org and it runs your program. |
| postpeer-pilot | A performance-driven publishing autopilot for short-form video, as an MCP server. 24 reliability/invariant tests, a 6-case agent-eval suite that passes 18/18 across 3 trials (incl. all three unsafe-action-refusal cases, 9/9 runs — the agent tried to schedule live, the tool layer refused), and a backtest that reports honestly where its own conservative design costs performance, not just where it wins. |
| meta-youtube-comment-mcp | Human-in-the-loop comment automation for Instagram, Facebook & YouTube. 46 deterministic tests, a written threat model, an eval harness that measures classifier + routing accuracy on labeled data — not just vibes. |
| know-no-hearsay | A reference architecture for evidence-bounded generative media: no pipeline step may increase the epistemic strength of a claim without evidence. snakeoil-radar (finds viral health-misinformation claims on TikTok, checks them against PubMed) and cited-cuts (turns a found claim into a quote-bounded rebuttal video) are two applications of it. |
| sorting-hat | An LLM-classified document inbox: drop scans into one folder, get named, filed paperwork out. Boring, useful, runs daily on my machine. |
| inanimatus | LLM-driven CAD, done declaratively — geometry-tested CadQuery models from dimensioned specs, as a documented method + Claude Code skill. |
📫 LinkedIn · 🎬 the daily-shipping pipeline in action: @dabigredbutton