Spoilr Intelligence
Thesis
A market intelligence pipeline over audience discussion. It ingests fan conversation across YouTube, web forums, and social platforms, structures sentiment and prediction signals, and feeds both the Spoilr game engine and standalone reporting.
Under the hood
Platform collectors feed a LangGraph classification pipeline; structured signals land in Postgres. Runs unattended as a scheduled daemon. Alerting covers lag, cost, and drift. First week of collection: 1,200+ posts from 85 monitored sources across 18 shows, 350+ flagged prediction relevant.
Decisions
- 01
Batch over streaming.
Scheduled collection cycles, thirty minutes to twice daily and tuned per platform, match how discussion actually moves, and a hard $15/day spend limit caps the worst case.
- 02
Classify with small models, escalate hard cases.
The expensive model sees only ambiguous items. Quality holds, and spend doesn't grow with volume.