Problem — A proper privacy audit is expert, manual, and expensive — exactly the work that doesn't get done.
A Claude Code skill that audits any site for GDPR + ePrivacy end to end: a deterministic Playwright scanner produces the facts, the model judges them against a validated JSON contract, findings gate on screenshot evidence, and a second command drafts the complaint for the right authority. MIT-licensed, nine EU DPAs seeded.
Problem — Structural data tests pass a customer called Mickey Mouse and a support ticket with an IBAN in it, and regex can't read sarcasm or negation.
Four demos that put a small judgment model (TypeSafe's Jev) inside tools data engineers already use: dbt tests written as English sentences, semantic SQL functions in DuckDB, a VISUALIZE clause that picks the chart, and a LinkedIn post grader. Code proposes, the model returns a typed judgment, code decides. Every number is scored against an answer key or baseline from a logged live run.
Problem — Your X/Twitter export is a useless pile of JSON, and the search you'd want would mean handing it all to someone else's server.
Parse the full export, generate embeddings locally with Transformers.js, search by meaning, and batch-delete with surgical precision — entirely on-device. A complete terminal UI in React/Ink with strict TypeScript and SQLite + FTS5 + sqlite-vec for hybrid lexical/vector search.
Problem — Picking a model for an agentic task on vibes is how you ship something slow, wrong, or expensive without knowing which.
Pit LLM providers against each other on real agentic tasks and measure what matters — correctness, latency, token usage, and cost — with structured, repeatable evaluation. CLI-first and CI-ready, so the benchmark runs the same every time.