Raw records receivedReal
—Returned by the public openFDA endpoint.
Public, inspectable proof
From market signal to guarded revenue action.
A production-style GTM system that ingests public FDA enforcement signals, scores accounts with a disclosed policy, and prepares evidence-grounded briefs for human review. Every number below is labelled by what it actually proves.
Loaded from the committed real-data fixture. Browser refresh is read-only and is not persisted.
Auditable run receipt
Evidence ledger
Live ingestion run: loading…
Signals persistedReal
—Passed the fail-closed identity guard.
Accounts scoredDerived
—Ranked with the disclosed openfda-v1 policy.
External writesGuarded
0Salesforce, HubSpot, and Clay remain preview-only.
System design
Production controls around the model
The LLM interprets evidence. It does not own truth, scoring, or delivery.
Strict tool schemas
Tool inputs and final outputs are validated before persistence.
Cost budget
Tokens, latency, and estimated spend are stored per run.
Retries + limits
429 and transient failures use bounded retry and backoff.
Evaluation gates
Valid citations are not enough: source meaning, score, identity, and fabricated contacts are checked.
Real public fixture · deterministic output
Account signal review
70+ priority 50–69 watch
| Account | Score | Signals | Latest | Class |
|---|
No accounts match this filter.
Structured output example
Agent brief
Qualification—
Human-review role—
Suggested angle
—
Why this is safe to inspect
The committed example came from deterministic mock mode: no Claude call, no fabricated contact, and no message sent.
Activation boundary
Integration preview
Payload mapping is implemented. Delivery requires explicit live switches and private credentials that are not committed.
Repository proof