Haute Fidélité · Automation
HiFi watch: the daily bot repaired, expanded (30 → 64 sources) and hardened
Context
The bot that sends a HiFi watch to Telegram every morning (the world's hi-fi news) had been down since June 15: the AI model that writes the summaries was hard-coded and had reached end of life. No brief was going out anymore. Beyond the repair, a chance to widen coverage and harden the whole thing.
What was done
- Repair. The AI model is now auto-selected — the watch can no longer break when the provider updates.
- Sources: 30 → 64. French specialist and mainstream press (ON-mag, AVCesar, Diapason, Vumètre, the Son-Vidéo blog, Les Numériques, Frandroid, 01net…), wider Europe (DE, IT, ES, NL, BE, SE, PL), Reddit, Head-Fi and diyAudio forums, reference YouTube channels.
- Brief format. A 🇫🇷 Made in France section first, then 🌍 Around the world. Full summaries in French (understand everything without opening the link), sources cited.
- Deduplication. A story picked up by several sources = a single block, with an importance signal.
- Reliability. HTML Telegram delivery (formatting that holds), clean splitting, automatic alert if a source dies 3 days in a row, ~1-month anti-duplicate memory.
- Cost under control. About €3-4/month, capped.
Validated in production
Last real run: 125 articles → 123 topics → 4 messages on Telegram, formatted, sent error-free. A slow source (LowBeats) was handled cleanly, without failing the brief.
Next steps (parked)
- Structured export of the watch to the content machine (
hl-content). - An "emoji → draft" flow to push an item toward a piece of content.
- Publishing on the site + a subscription offer (to be scoped).
Stack
Python, automatic AI model selection (API), Telegram API (HTML delivery), deduplication with ~1-month memory, a daily cron with a dead-source alert. Repo veille-hifi.
What this illustrates
A hard-coded dependency (here the AI model) is a time bomb: the day it disappears, everything stops silently. Making it auto-selected removes the entire class of failures. And for a watch tool, reliability (deduplication, dead-source alert, capped cost) matters as much as the number of sources: a brief that doesn't arrive, or repeats itself, helps no one.