systemsts2026source
instagram listening for a small brand: every comment and mention, tagged for sentiment and intent, in roman urdu too
end goal → see every mention and tag of the brand in one place, and answer each one in a way that helps it
The problem
A small brand's Instagram is where people complain, ask and recommend, and none of it waits for office hours. The comments that matter sink under the ones that do not, and a lot of them arrive in Roman Urdu, which most sentiment tools read as noise.
What I built
Petal takes in every comment, @mention and tracked hashtag post, tags each one for sentiment and intent, and shows it as a live feed, seven days of stats and a set of alerts.
- Webhooks and polling, both. Meta delivers webhooks on a best-effort basis, so a poller backfills anything missed. Both write through one upsert keyed on Instagram's own ids, which makes a duplicate or a retry harmless.
- AI never sits in the ingest path. A BullMQ queue sits between ingestion and the model, so a slow or failing LLM call cannot drop a comment.
- Claude Haiku with structured output, with repeat text served from a sha256 cache. A daily budget hard-stops spend, and anything over budget or failing falls back to a keyword list that understands Roman Urdu.
- Daily rollups, not scans. Stats read from precomputed per-day totals. The trade is a few seconds of lag for a dashboard that stays fast as the history grows.
Where it stands
Every work package is done, and the demo above runs on seeded data, because the account it was built for is new and quiet. What remains needs other people's permissions: Meta's app review, a second worker service and a real key, so the AI's accuracy can be published rather than claimed.
all projects