in production since June 2026A Telegram bot that finds jobs for me. It pulls postings from 18 sources, drops the wrong ones with simple rules before any model call, and has a model score the rest.
37,702
postings collected by October 2026
47 of 60
good postings (ones I applied to or the bot sent me) pass the scorer I shipped. The old one, re-run, passes 35.
One thing I caught by measuring. The bot was paying a model to score postings that had nothing to do with my work, nurses and lawyers among them: 1,365 of 4,845 paid scores in three weeks came out near zero. I added a free check on the job title that runs before any paid call.
More about GigaWork
- I test scoring on 80 real postings labelled by outcome. After a model change I re-ran them: scores fell about 8 points and the live threshold of 60 still held.
- Posting text reaches the model wrapped as data, as a defence against prompt injection. The agents that read it get only allow-listed tools.
- The job queue is plain Postgres. Workers take rows with an atomic claim, with no separate queue service.
- It also built CVs and cover letters on demand from a structured record of my whole career, with the facts pinned so nothing got invented.
- If one model provider fails, the next one takes over. A failed provider is parked for its retry window (a circuit breaker).
- 11 LLM-judge agents check each generated CV, and two can block it. The judges are not yet checked against human labels.
- An e-mail application goes out only after I press Send, and never twice.
- I measured the Batch API at 2.3 times the price of cached calls ($0.0047 against $0.0020 per posting), so I dropped it before launch.
TypeScriptNext.jsNode.jsPostgres (Neon)SQLDrizzleClaudeGeminiZodVitest
Mixing a multitrack recording in the browser. You type what is wrong ("the vocal is buried") and a model turns it into a checked list of changes.
- A cheap model answers first. If the answer fails validation or the call fails, it retries once on a stronger model.
- Guardrails live in code, not in the prompt: the server clamps every change the model proposes to an allowed range.
More about GigaMixer
- A test set of 150 labelled phrases with pass thresholds, including attempts to trick the model.
- A Python worker separates instruments from a recording with local models.
Next.jsAnthropic SDKtool callingshadcn/uiDemucsSupabaseCloudflare R2
The site of my recording studio. Artists leave comments right on the audio waveform, and I update the site from a Telegram bot. It has its own design system.
- 63 design tokens (colours, fonts, spacing) from one JSON source, generated as code for web, iOS, Android and Windows.
- The build fails if text contrast drops below WCAG AA, the accessibility standard.
Next.js 16React 19Server ComponentsWeb Audio APIdesign tokensDrizzle
Picks one new release a day from my Bandcamp collection. A small audio model compares how releases sound, and that raised ranking quality (MRR) from 0.118 to 0.200. I tuned and tested on the same data, so the number is optimistic.
TypeScriptNode.js 20ONNXembeddingsGitHub Actions
More projects
quizMy study tool. A model grades spoken answers about code against a rubric.
private repo
tg-summary-botSummarises long Telegram chats with Gemini. Python.
runs locally