Autonomous Ops Monitoring
Cut an automated report pipeline from 145 to 300s (failing about 1 in 3 runs) to a reliable 34 to 52s by moving unbounded reasoning out of the LLM.
The problem
Built a live monitoring dashboard for a multi-store market rollout: a daily data pipeline, automated anomaly flags, a durable issue tracker distinct from the auto-clearing flags, and an executive summary. The interesting part is not the dashboard. It is the specific failure this project diagnosed in how an LLM was being used inside an automated pipeline. The summary step was asking an LLM to re-derive, from scratch, on every run, which reporting day was the last complete one and a four-way status classification for every store, by scanning the full dataset each time. That unbounded reasoning task was bolted onto a job that needed to run reliably on a timer, and regeneration was timing out on roughly two of every three runs.
The fix
Moved the classification logic out of the LLM entirely and into deterministic Python, reusing logic that was already proven correct elsewhere in the same system’s front-end code. The LLM’s job shrank from "figure out what happened" to "write a paragraph about a pre-computed answer." Smaller, bounded input, and a job an LLM is actually well-suited to.
Stack
Python data pipeline, scheduled local automation (launchd), a headless Claude Code call scoped to prose synthesis only, HTML/JS dashboard front end.