Genomic Inference Rx Agent — pharmacogenomic clinical decision support for Type 2 diabetes. Combines 23andMe-style genotypes, clinician intake, synthetic CGM/WHOOP signals, live ClinVar/PubMed/ClinicalTrials.gov lookups, and an agentic pipeline that produces a structured clinician brief.

Frontend: Next.js provider UI in web/ (login, patient roster, Setup uploads, Brief, Metrics, Intake, Genome tabs).

cp .env.example .env
# Set NVIDIA_API_KEY (recommended) and NCBI_EMAIL

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# Demo data (optional)
python scripts/build_patient_file.py
python scripts/build_whoop_data.py
python scripts/build_glucose_data.py
python scripts/seed_db.py

uvicorn server.main:app --reload --port 8000
# OpenAPI docs: http://127.0.0.1:8000/docs

cd web && npm install && npm run dev
# UI: http://127.0.0.1:3000  (set NEXT_PUBLIC_API_URL in web/.env.local)

Verify the agent:

python scripts/test_agent.py PT-002

Agent brief response (JSON): safety_flags, snp_summary, recommendation, glucose_insight, wearable_insight, trial_matches, trial_search_meta, citations, intake_summary, patient_summary, _trace (each step may include reason and agent_role), _backend, optional _llm_model (when Nemotron ran on NIM/OpenRouter/Ollama).

Brief generation streams live tool-call events immediately; local demo runs are bounded by AGENT_TIMEOUT_SEC. Parallel mode brief assembly is capped by BRIEF_ASSEMBLY_TIMEOUT_SEC; if assembly times out, the partial brief still includes trial_matches merged from fetch_trials so the UI can list trials.

ClinicalTrials.gov (apis/clinical_trials.py): recruiting T2D studies are queried with a larger page size; US 5-digit ZIP codes are not sent as query.locn (the v2 API often returns zero hits for bare ZIPs). Search is national with optional distance ranking when the ZIP is known. Parallel runs dedupe trials by nct_id / nctId.

Clinician loading UI (web/components/gira/gira-loading-page.tsx): full-screen pipeline shows a DNA helix, progress from trace steps plus a time ramp, and an Agent activity column on the right where completed steps reveal one-by-one (~680ms apart) while the run is active; everything flushes when the run finishes or is cancelled. The Agent tab passes the streamed brief to the Clinician brief tab via a ref so trial_matches is not lost to a one-frame React lag.

Structured intake lives in schemas/patient_intake.py and is stored as JSON on the patient row (intake_json) or under data/intake/{patient_id}.json.

Fields: medications (name, dose, frequency), vitals, goals, sideEffects, lifestyle, comorbidities, familyHistory, clinicianNotes.

The agent loads intake at the start of every run and uses it for med matching, goals, and PGx synthesis context.

GIRA/
  server/           # FastAPI app
  agent/            # tools.py, memory.py
  reasoning/        # nemotron loop, safety_flags, pgx, prompts
  output/           # assemble_brief (generate_brief tool)
  parsers/          # genome, WHOOP, glucose, intake
  apis/             # ClinVar, PubMed, trials, RxNorm, PharmGKB stub
  schemas/          # intake form types
  scripts/          # seed, reset, test_agent, data builders
  data/             # synthetic genomes, whoop, glucose, intake (no PHI)

GIRA is positioned as an autonomous clinical PGx agent: Nemotron (or deterministic fallback) drives a tool loop with live APIs. Defaults in .env.example match submission guidance:

• AGENT_MODE=llm — Nemotron-visible ReAct loop (run_with_tools in reasoning/nemotron.py).
• PGX_SYNTHESIS=1 — optional writer pass over SNP text + citations from ClinVar/PubMed only (reasoning/pgx_synthesis.py).

Demo patient: PT-003 — SLCO1B1 + CYP2C19 safety, CPIC, and citations (see scripts/demo_hackathon.sh).

Devpost-ready copy: DEVPOST.md.

Models: NIM NIM_MODEL (default nvidia/llama-3.3-nemotron-super-49b-v1.5) or OpenRouter OPENROUTER_MODEL / OPENROUTER_AGENT_MODEL (default Super 120B free slug). Override OPENROUTER_AGENT_MODEL if the default is throttled.

After seed_db.py, intake JSON is loaded from data/intake/PT-00X.json.

Deterministic checks in reasoning/safety_flags.py (always run before brief assembly):

• SLCO1B1 rs4149056 TT → statin myopathy risk
• CYP2C19 rs4244285 AA → clopidogrel poor metabolizer
• VKORC1 rs9923231 AA → warfarin hypersensitivity
• SLC22A1 rs622342 AA → metformin reduced efficacy

See .env.example. Minimum for production-like runs:

• AGENT_MODE=llm and PGX_SYNTHESIS=1 — Nemotron-visible loop + optional synthesis (defaults in .env.example)
• NVIDIA_API_KEY — LLM backend (Nemotron on NIM)
• NCBI_EMAIL — PubMed / ClinVar (optional NCBI_API_KEY for higher rate limits)
• USE_SYNTHETIC_WHOOP=true, USE_SYNTHETIC_GLUCOSE=true — demo wearable data

Optional tuning:

• AGENT_BRIEF_CACHE — when true, GET /agent_brief can return a persisted SQLite brief without re-running tools.
• BRIEF_ASSEMBLY_TIMEOUT_SEC — parallel mode cap for generate_brief (raise if PGx synthesis often times out).
• AGENT_TIMEOUT_SEC / AGENT_TOOL_TIMEOUT_SEC — overall and per-tool demo ceilings.

Research and demonstration only. Not a substitute for licensed clinical judgment. All outputs require physician review.