🧭Message Router

HackerRank Orchestrate · Ranked 18 / 1,983

Every WhatsApp message,
routed like a person would.

An AI pipeline that decides what interrupts you now, what waits for later, and what gets muted — reasoning over text, image posters and voice notes, with hard safety rules a language model can't be talked out of. This isn't a mockup: every verdict below is a real, precomputed output of the actual routing run. Tap in.

93.3%action accuracy
32 / 110API calls for messages
$0cost, free tier only
15false claims self-caught
Tap any dashed message → see the full 5-phase trace↳ real signals, not vibes
Switch inboxes — same sender, different verdict, on purpose↳ personalization is the whole point
9:41 📶 🔋

How a message actually gets routed

Five phases, most of them deterministic. The model only speaks once it's the last honest option.

📄 dataset/*.csv 1 · bundle.py one block per user 2 · media.py OCR + transcript, cached 3 · rules.py 4 hard overrides override fires? verdict skips the LLM 4 · router.py 1 call / user, T=0 5 · normalize.py semantic check 📤 output.csv override's own evidence still reaches router.py
Rules can force the verdict on 4 hard-override conditions (OTP/credential asks, lookalike-domain payments, router-injection attempts, chain forwards) — but the router's evidence selection survives the override, since evidence is scored output, not a safety decision.