Inspection-ready pharmacovigilance AI

The pain we built this for

If you run a drug-safety or pharmacovigilance operation, the math is against you: “Case volume climbs 20–30% a year, every ICSR is a regulatory clock, our narrative quality is reviewer-dependent, and the one thing we cannot ship is a hallucinated adverse-event fact — so we can’t just point an LLM at it.”

• ICSR volume grows every year while headcount does not — intake, triage, coding, and narrative writing all compete for the same reviewers against the same regulatory clock.
• A generic LLM is disqualified on day one: a single hallucinated causality or seriousness statement in a safety narrative is not a bug, it is a compliance event.
• Signals hide in the case series. By the time manual triage clusters the related reports, the pattern is weeks old.
• 2026 FDA and EMA guidance expects AI in safety workflows to be explainable, validated, human-in-the-loop, and inspection-ready — not a black box that files on its own.
• You cannot send patient narratives or proprietary signal data to a third-party AI API — so most "AI PV" offerings are off the table before the first demo.

See it — a drafting agent that can’t file, on a ledger that can’t be altered

This is the whole control model in a handful of SQL statements: an agent bounded to propose, a source-grounded draft, and an encrypted, tamper-evident audit ledger an inspector can verify.

-- A drug-safety drafting agent that PROPOSES but never files.
CREATE AGENT pv_case_drafter
  PERSONA 'You are a pharmacovigilance case processor. Draft the ICSR
           narrative ONLY from the cited source-case fields. If a fact is
           not in the case, say so — never invent it. Propose seriousness
           and causality for a qualified reviewer to confirm.'
  ALLOW_WRITES FALSE;                 -- proposes; a human approves

-- The inspection-ready audit ledger: tamper-evident + encrypted at rest.
CREATE IMMUTABLE TABLE pv_audit (
  case_id      TEXT,
  drafted_by   TEXT,
  model_version TEXT,
  sources      JSONB,                 -- exactly which case fields grounded it
  draft        TEXT,
  reviewed_by  TEXT,
  decision     TEXT                   -- approved / edited / rejected
) WITH (ENCRYPTION = 'AES256GCM');

-- Draft one case, grounded in its own record.
SELECT AGENT_RUN(
  'pv_case_drafter',
  'Draft the ICSR narrative for case ' || :case_id ||
  ' using only the reporter narrative, dechallenge/rechallenge, and
    concomitant medications on file. Cite each field you use.'
);

-- Later, for the inspector: prove the trail was never altered.
VERIFY TABLE pv_audit;

Seven safety workflows. One self-hosted stack.

Prove what the agent drafted, prove the record wasn’t altered, and prove nothing left your environment — in one engine, not a stitched-together pipeline.

The objections a safety lead raises first

Data residency, AI trust, validation, and footprint — answered before you have to ask.