Attention Router
Notification triage that learns what each recipient actually does.
Python · FastAPI · BM25 · LLMs
- Classifies messages across three modalities — text, images and voice — into notify, digest or mute, using recipient-specific behavioral history and tiered BM25 retrieval.
- A deterministic safety layer catches phishing, brand impersonation, credential requests and prompt-injection attempts before the language model is ever invoked.
- Production-ready LLM pipeline: schema-validated outputs, citation extraction, retry and fallback handling, provider-pluggable models and offline tests in CI.