Small tools, each built because something kept going wrong

I use these daily. Most of them exist because a check I trusted let a mistake through.

mail_check.py

Linter for LLM-drafted email. Rules come from a written retro of every rejected draft. Exit code 1 blocks the draft.

web_check.py

Website QA: horizontal overflow and overlapping elements at several screen widths, contrast, console errors, failed requests, broken links.

prejimka.py

Website audit that reports only findings with hard proof: a cookie name, an HTTP code, exact console text, a selector. No heuristics, no taste.

jev.py

Structured classification: yes/no, pick one, or a scale. Returns probabilities, not prose. Batch over CSV, logs confidence because Czech is weaker than English.

nv.py

Model router for routine work with provider fallback. Free endpoints log prompts, so client content never goes there.

recall.py

Semantic search over my notes (~5,600 files) with incremental indexing, so I find what I already decided before starting again.

leadguard.py

Contact ledger with identity by company ID, then domain, then email. Filter before drafting, mark after sending.

cz_norm.py / cz_denorm.py

Czech text to speech-ready words and back. Dictated numbers become digits by code, because the LLM got a phone number wrong.

asr_score.py / asr_bench.py

Transcription with word error rate, and a blind test of human versus machine on any audio.