I find the thing that does not match, prove it, and ship the tool in the open.
I spent 12 years as an intelligence analyst in the Pennsylvania Army National Guard, including a deployment to Kosovo as a counterintelligence analyst. Today I work in life-safety monitoring and run a small research and AI search (GEO/AEO) practice at ryanlenk.com.
What I study: how ranking systems and AI answer engines decide what to surface and what to trust, and how people game them.
| Project | What it does | What's in the repo |
|---|---|---|
| candor | Twelve anti-sycophancy work-mode personas for Claude Code: coding, debug, security, architecture, writing, data and more. | Blind A/B evals. Candor won 14 of 20 pairs (baseline 4, ties 2). Prompts, scoring scripts and the losses are committed. |
| humanist | A prose checker and an offline sanitizer for AI-assisted writing. | False-positive sweep on 24 known-human documents from 1854 to 2018, with its limits stated up front. |
| bangermeter | Scores any post with X's open-sourced heavy-ranker formula and last published weights. Chrome extension and userscript. | On the Chrome Web Store. Engine self-test included. |
| cutlist | One Agent Skill: a long video and its captions in; a clip plan, titles, descriptions and one thumbnail out. Works in Claude Code, Codex, Cursor, Gemini CLI and Copilot. | Never uploads, posts or logs in to anything. |
| assay | Local-first desktop tool that finds local businesses, checks their page source with six deterministic checks, and drafts an outreach packet. | No sender in the codebase. Every artifact needs its own approval. |
| lenkraster | Pixel-art critique, palette and animation QA, with an Aseprite bridge. | Local MCP server with seven tools. |
| wainwright | Skills that teach an orchestrator bot to design one-job Grok Bot teammates and lint them before they ship. Cursor plugin. | Experimental, v0.1.0. |
Also live: Dino Disasters, a free co-op dino arena game for 1 to 4 players that runs in the browser, and Galatéa, a weekly AI news and tool-review site written by an AI persona.
- Every claim gets a confidence grade, from proven to unproven.
- Results that went the wrong way get published too.
- Local-first. Nothing sends, posts or publishes without a human click.
- MIT unless a file says otherwise.
- Leading a team in the Muse Spark Open Kernel Challenge, which runs through Nov. 19, 2026.
- Writing at ryanlenk.com.
ryanlenk.com · hello@ryanlenk.com · @itsryanlenk on X · LinkedIn


