Anukriti — Pre-Prescription Safety Agent

A Gemini agent that catches ancestry-specific drug risks before a prescription — deterministic CPIC checks, named rules, and live Elasticsearch-grounded evidence.

Comment

Inspiration

In 2019, a relative — South Asian, post-stent — was prescribed clopidogrel.
He carries CYP2C19 2/*2: a Poor Metabolizer who literally cannot activate
the drug. It was, for him, a placebo against stent thrombosis. He had a
second cardiac event. **14% of South Asians can't activate clopidogrel, yet
they're prescribed it at the same rate as Europeans, where ~2% are affected.*
Pharmacogenomic guidelines are calibrated on Eurocentric evidence, and the
gap is invisible at the point of care. Anukriti is our answer: an agent that
asks "will this drug work for this person, given their ancestry?" — before
the prescription is written.

What it does

Anukriti is a pre-prescription pharmacogenomic safety agent. You ask about a
family member's medication or upcoming procedure, and it:

  • Looks up the person's ancestry + medications (MongoDB Atlas)
  • Runs a deterministic CPIC-pinned safety check — never an LLM guess
  • Returns a named rule ID for every conclusion (R12 SUFFICIENT, V10 SUPPORTED,
    U9 LOW uncertainty) and an honest named refusal when evidence is thin
  • Grounds every recommendation in live evidence retrieved from
    Elasticsearch     (CPIC guidelines, PubMed, PharmGKB)
  • Surfaces ancestry-specific risks standard tools miss — most strikingly the
    BCHE L307P variant, carried by ~5.3% of the Vysya community (≈18× the
    all-India rate, absent from gnomAD), which causes succinylcholine paralysis
    during anesthesia

The flagship demo: a grandmother traveling to Hyderabad for cataract surgery.
The agent catches both her ineffective clopidogrel and a $30
butyrylcholinesterase test that prevents a multi-hour anesthesia emergency —
a risk no commercial PGx test screens for.

The LLM explains; it never decides. Every claim cites a rule; every rule
cites a paper; every refusal is named.

How we built it

  • Agent: Gemini (Google Cloud Agent Studio / Gemini Enterprise Agent
    Platform), with a hardened system prompt that forces grounding-first and
    rule-ID narration.
  • Tools: 7 MCP tools served from a Cloud Run container, exposed to
    Agent Studio via a Streamable-HTTP MCP server endpoint (/mcp).
  • Deterministic engine: anukriti-pgx-core (CPIC 2022 pinned tables),
    composed in-process behind the tools — no LLM in the calling path.
  • Elasticsearch (partner track): a pgx-evidence index seeded with the
    exact papers the engine cites; the search_pgx_evidence tool runs a live
    hybrid query so every citation resolves to a real, retrievable source.
  • MongoDB Atlas: per-user family rosters, medications, travel plans, runs.

Challenges we ran into

  • Agent Studio attaches custom tools only via an MCP server (no OpenAPI
    import, no-auth only). Our tools were REST, so we wrote a thin
    Streamable-HTTP MCP wrapper over the existing functions.
  • Smaller models occasionally guessed an ID instead of grounding first; we
    fixed it with an explicit "always call list_family_members first, never
    invent an ID" hard rule.
  • Validating partner SDKs against their current contracts surfaced three
    real breaks (Elasticsearch body= deprecation, Fivetran endpoint rename,
    Arize v8 API) — caught before they shipped.

Accomplishments that we're proud of

  • A medical-safety agent that is hallucination-resistant by construction:
    deterministic rules first, LLM only narrates, every refusal named.
  • Live, load-bearing Elasticsearch grounding — the agent's evidence step
    is a real hybrid query, not a static list.
  • The BCHE/Vysya wedge: surfacing a 2025 Cell-published variant that
    gnomAD-based commercial tests are blind to.
  • 57 passing tests; the engine is independently validated against an external
    warfarin cohort (n=5,700).

What we learned

Equity in drug safety is an infrastructure problem, not a model problem.
A bigger LLM over Eurocentric evidence just produces confident wrong
answers. Treating evidence density as first-class — and refusing honestly
when it's thin — is the real fix.

What's next for Anukriti

Wire the remaining partner integrations live (Dynatrace self-observability,
Fivetran cohort ingestion, Arize narration eval — all built), expand beyond
the 3 CPIC workflows, and pursue institutional data partnerships for
controlled-access ancestry data.

Built With

Share this project:

Updates