Email me to unlock the full case study and demo.
News ingestion → classification → sentiment drift monitoring — 10k docs/hr at p95 < 300ms.
Feeds (RSS, APIs, social firehose) are de-duplicated and normalized. A relevance classifier filters finance-specific content, then sentiment is scored with a hybrid approach: fast prompt classifier for bulk, and a finetuned transformer for edge cases. Metrics are tracked and drift signals trigger an active learning loop with human-in-the-loop review.
Deployment is containerized with autoscaling workers. The pipeline exposes a FastAPI/WS interface used by downstream models (like gold forecasting) and the dashboard UI. Alerts are sent when latency or quality falls out of bounds.