The spine of the entire system. Every agent reads and writes one case object — nothing lives outside it. 8-table PostgreSQL schema: patients, prescribers, payers, drugs, pa_cases, case_requirements, case_documents, case_communications, case_events.

All transitions enforced through a single transition(case, to_status, actor) service function that writes case_events. Agents may propose transitions to ready_to_submit, submitted, and appeal_submitted — but only a human actor may execute them.

pgvector pipeline over payer PA criteria documents — Caremark, Express Scripts/Evernorth, OptumRx, Aetna, UHC, state Medicaid PDLs. Chunked per-criterion, not per-page. Hard filter on payer + drug_generic + effective date window, then semantic rank.

Reads electronic prescriptions, faxes, paper, and phone orders. Uses OCR, Vision models, and LLM extraction to capture drug, strength, quantity, directions, prescriber, and insurance details. Creates the PACase in intake status.

First intelligent decision point. Checks NCPDP reject codes: 75 (PA Required), 76 (Plan Limitations Exceeded), 3W/MR (PA in process), 70 (Product Not Covered → formulary alternative path), 88 (DUR reject → different workflow).

Automatically checks BIN, PCN, Group, Member ID, coverage status, and preferred drug. Detects step therapy requirements, quantity limits, age restrictions, and diagnosis restrictions from payer data.

Derives requirements ONLY from provided criteria excerpts. Cites chunk_id for every requirement. Splits compound criteria into individually checkable items. Classifies each as diagnosis / lab / step_therapy / quantity / age / other.

Marks each requirement satisfied only if it can point to specific evidence (document + location + value). Quotes evidence into evidence_note. Ambiguous or stale evidence = NOT satisfied. denial_risk reflects the weakest link, not the average.

Generates medical necessity letter structured as: patient/drug header → diagnosis with ICD-10 → prior therapies with dates and outcomes → current clinical status with lab values → why this drug meets the payer's stated criteria, mirroring the payer's own criteria language → prescriber attestation block.

Extracts the payer's stated reason verbatim before classifying it. Extracts appeal deadline and appeal channel exactly as stated. Distinguishes "denial is correct given submitted evidence" from "denial is wrong" — the appeal strategy differs completely.

Addresses the specific denial reason head-on in the first paragraph — an appeal that restates the original request loses. Cites clinical guidelines by name (ADA Standards of Care, NCCN) only if confident, otherwise inserts [VERIFY]. Includes a documents-enclosed list.

Generates outbound request from gap analysis. One generator, three renderings: fax cover sheet (PDF), phone script, portal-message text. Asks only for what the gap analysis marked missing — offices ignore kitchen-sink requests.

Deliberately not an LLM — a Celery beat job running every 15 minutes. Escalation ladders defined as data, not code. Deadline override: if payer_deadline or appeal_deadline is within 48h, case jumps the ladder and goes straight to pharmacist escalation.

Twilio SMS + voice with BAA. Two hard rules: (1) No PHI in message bodies — not the drug name, not the condition. (2) sms_consent must be true before the first message; first message includes opt-out language (TCPA requirement).

Manual-first monitoring until Phase 3 lands. Kanban UI over pa_cases.status — columns = statuses, cards sorted by next_action_at, one-click status updates that fire the state machine. Aging badges: yellow ≥3 days in column, red ≥ deadline−48h.

Fax received → vendor webhook → download PDF → S3 (SSE-KMS) → Textract OCR → fast-model classifier → auto-attach to matching case. Confident match: notify. Ambiguous: route to triage queue for human to attach.

Parses nightly PMS export of rejected claims. NCPDP codes that matter: 75 (PA Required), 76 (Plan Limitations Exceeded), 3W/MR (PA in process), 70 (Product Not Covered), 88 (DUR reject). Auto-creates pa_cases in intake status.

SQL over pa_cases + case_events + case_communications. Core views: v_stage_durations (cycle time per stage), v_headline (open cases, avg days to fill, approval rate, revenue recovered MTD). No BI tool needed.

Formulary-alternative suggester (no case volume needed — build early). Pre-submission denial predictor (after ~100 decided cases). Prescriber-latency-aware follow-up. Worklist prioritization: estimated_revenue × approval_likelihood ÷ days_to_deadline.