AI operations digest agent
An agent that watches operational sources, removes duplicates, summarizes changes with evidence, and delivers one daily brief.
- Role
- AI automation engineer (demo)
- Timeline
- 4-week prototype (demo)
- Stack
- TypeScript · LLM / Ollama · Linux systemd timers · Postgres · Drizzle · RSS + HTTP
Demo case study — figures illustrative; client redacted.
What was broken
An operations team checked release feeds, status pages, repositories, and internal notices by hand. Important changes were mixed with repeats and low-signal updates, while copying everything into a general-purpose chatbot lost source history and delivery guarantees.
How it was fixed
Built source-specific watchers with durable cursors in Postgres, normalized and deduplicated each event, then passed only new evidence to an LLM for a structured impact summary. A systemd timer runs the pipeline, records every stage, retries failed sources independently, and delivers a daily digest with links back to the original evidence.
Constraints
- Preserve a source link and collection timestamp for every summarized item.
- Deliver on schedule even when one upstream source is unavailable.
- Run on a modest Linux host with a replaceable LLM provider.
Tools in the system
- TypeScript
- LLM / Ollama
- Linux systemd timers
- Postgres
- Drizzle
- RSS + HTTP
What changed
Replaced scattered source checks with one reviewable daily artifact while preserving provenance and isolating collection failures from delivery. (Figures illustrative — demo case study.)
What carries forward
- 01The agent needs durable state before it needs better prompts; cursors and idempotency prevent duplicate briefs.
- 02Summaries stay trustworthy when every item carries evidence and the model is forbidden to fill missing context.
- 03Source failures should degrade one section of the digest, not block the entire scheduled delivery.
Related case studies
All work- AI & Agents · 2026Self-hosted GPU inference platform — 36B MoE, voice AI and a 70-GPU farmA fully self-hosted AI stack: a 36B MoE model served concurrently with embeddings, ASR and TTS on a single 24 GB GPU — later scaled to a 70-GPU, 10-node farm that a bad VBIOS clock state nearly took down.
- AI & Agents · 2026OCR document analyzer — waste-manifest ticket intakeA vision-LLM system that reads ~50 WhatsApp scale-ticket scans a day and replaces manual re-typing with validated extraction and one-click export.