PV Sentinel

An AI-assisted pharmacovigilance platform built patient-first and audit-ready.

Draft regulator-ready adverse-event narratives faster ΓÇö without losing the patient's voice or the audit trail.

---

Status: Working MVP, built 2025. Feature-complete against the scope below; not validated for regulated production use (see Scope and next steps).

About this repository: this is the public overview of PV Sentinel. The application source is held in a private repository, which carries the full development history. This overview is generated from the codebase, so the capabilities described here track what is actually built.

Built by: Wayne Kearns ΓÇö designed and built end to end.

---

The problem

Pharmacovigilance (PV) ΓÇö the science of monitoring drug safety ΓÇö runs on a slow, manual, and error-prone workflow:

• Writing a single adverse-event (AE) case narrative can take 2ΓÇô4 hours.
• Manual medical coding is inconsistent and error-prone.
• Established enterprise safety systems are expensive and require long, heavy deployments.
• Regulators (FDA, EMA, PMDA, Health Canada) demand complete, tamper-evident audit trails and reproducible records.

The result: safety teams spend disproportionate time on administrative drafting rather than on the medical judgement that actually protects patients.

What PV Sentinel does

PV Sentinel helps drug-safety professionals move from raw report to submission-ready case faster, with an AI drafting assistant that augments ΓÇö never replaces ΓÇö human review:

1. Intake structured, ICH E2B-aligned case data — patient demographics, suspect product and therapy dates, event onset, seriousness criteria, dechallenge/rechallenge, and the patient's own account.
2. Draft an ICH E2B-aligned narrative with an AI assistant.
3. Preserve the patient's voice and check that the AI hasn't summarised it away.
4. Review ΓÇö a qualified human must approve or reject every AI draft.
5. Audit ΓÇö every action is written to a tamper-evident, append-only trail.

What makes it different

Most tools in this space optimise for feature breadth. PV Sentinel optimises for the two things that actually matter to patients and regulators: trust and traceability.

Principle	What it means in practice
🗣️ Patient-voice preservation	The patient's own words are stored verbatim and immutably. Every AI narrative is checked for how much of that original voice survived — low fidelity forces human review.
🔒 Reproducible AI	Every AI generation is locked to a cryptographic fingerprint of the exact model and prompt used, so any output can be reproduced and explained later.
🧾 Tamper-evident audit trail	Actions are recorded in an append-only, hash-chained ledger; altering any historical record is detectable — built with 21 CFR Part 11 expectations in mind.
🤝 Human-in-the-loop by design	AI never auto-approves anything. Narratives are always gated behind a named reviewer.
🎯 Honest AI	No fabricated "accuracy" or "confidence" scores. If the AI can't help, it says so clearly rather than inventing content.
🔓 No vendor lock-in	Designed for portable deployment (cloud or on-premises) with pluggable authentication.

Inside the review workspace

Every AI draft lands in a review screen designed so a qualified reviewer can trust it at a glance ΓÇö and prove that trust later:

• Side-by-side change view ΓÇö see exactly what the AI drafted versus what the reviewer edited before approval.
• The patient's own words, in view ΓÇö the verbatim account sits alongside the draft, with automatic alerts when too little of the patient's voice survives.
• Reproducibility panel ΓÇö the exact model and prompt fingerprint behind each narrative, so any output can be explained and reproduced.
• Per-case history ΓÇö the complete, hash-chained trail of everything that happened to that case, from intake through review.
• Export the reviewed narrative — copy or download the final narrative as plain text once the case is done.

See it in practice

	Review workspace. The AI draft with the reviewer's edits shown inline, alongside the patient's verbatim account and a live patient-voice fidelity check.
	Reproducibility panel. The model and prompt fingerprint behind this narrative.
	Case history. The hash-chained record of every action taken on the case.

Synthetic case data only ΓÇö nothing here is a real adverse-event report or a real patient's words.

Who it's for

• PV / Safety Officers & Medical Reviewers ΓÇö faster drafting, consistent structure.
• Regulatory Affairs ΓÇö submission-ready records and complete audit trails.
• Quality & Validation ΓÇö reproducibility and inspection-readiness.
• Data Privacy / Governance ΓÇö patient data handled with care by design.

Architecture at a glance

PV Sentinel is a modern web application with a clear separation between a Python API (where the safety-critical logic and AI live) and a TypeScript web client.

flowchart LR
    U[Safety professional] --> FE[Web client<br/>React + TypeScript]
    FE -->|REST API| BE[Application API<br/>FastAPI / Python]
    BE --> DB[(PostgreSQL)]
    BE --> AI{{Pluggable AI provider}}
    BE --> AUD[[Hash-chained<br/>audit ledger]]
    subgraph Safety core
      PV[Patient-voice<br/>preservation]
      MT[Model & prompt<br/>version locking]
    end
    BE --- PV
    BE --- MT

Loading

Technology stack

Layer	Technology
Frontend	React + TypeScript (Vite)
Backend / API	FastAPI (Python), SQLAlchemy
Database	PostgreSQL (production), SQLite (local dev)
AI	Provider-agnostic large language models (initially Anthropic Claude)
Compliance foundations	Hash-chained audit trail, model/prompt hash-locking, role-based access control

Design principles

• Patient safety is a hard constraint, not a feature flag.
• Every AI output must be reproducible and reviewable.
• Scope discipline: ship a validated core before breadth.
• Secrets never live in the codebase; auth is pluggable and portable.

Regulatory posture

PV Sentinel is designed for an environment where every output has to survive inspection. The compliance foundations are built in rather than added later:

• ICH E2B alignment. Narratives are structured against E2B expectations so case data maps cleanly to downstream submission formats.
• 21 CFR Part 11 principles. The audit ledger is append-only and hash-chained, so any alteration of a historical record is detectable. Actions are attributed to a named, authenticated user.
• Reproducibility. Each AI generation is locked to a cryptographic fingerprint of the exact model and prompt used, so any narrative can be reproduced and explained after the fact.
• Human oversight as a hard constraint. No AI output reaches a record without approval by a qualified reviewer. This is enforced in the workflow, not left to policy.
• Data protection. Patient data is held in the application's own database, stored immutably and integrity-hashed, with access restricted to authenticated, role-scoped users. Retention policy and pseudonymisation are defined at deployment time.
• Accessibility. The interface is developed against WCAG 2.1 AA, with automated conformance checks (axe-core) gating every change in CI ΓÇö relevant to obligations such as the European Accessibility Act and Section 508. (Full conformance also requires manual assistive-technology testing.)

Formal computer system validation (IQ/OQ/PQ) would be required before use in a regulated production environment. See Scope and next steps.

Scope and next steps

PV Sentinel was built to prove a specific thesis: that AI can take real time out of adverse-event narrative drafting without weakening the audit trail or displacing human judgement. Scope was deliberately held narrow to get that core right.

Built

• Structured, ICH E2B-aligned AE case intake — patient, suspect-product and therapy detail, event and causality (dechallenge/rechallenge), seriousness criteria, reporter, and the patient's verbatim account
• AI-assisted drafting of ICH E2B-aligned narratives (provider-agnostic; honest failure when no model is configured)
• Patient-voice fidelity checking, with low fidelity forcing human review
• Reviewer workspace with side-by-side draft-versus-edit comparison and inline patient voice
• Model and prompt fingerprinting for reproducible generation
• Hash-chained, append-only audit ledger with integrity verification
• Role-based access control
• Case list with status filtering and free-text search
• Export of the reviewed narrative as plain text
• Versioned database migrations and a containerised deployment baseline (Docker + PostgreSQL)

Deliberately out of scope for the MVP

• Signal detection and case-series analytics
• Automated MedDRA coding
• E2B(R3) transmission to regulatory gateways
• Duplicate detection and case merging
• Multi-language intake

What production readiness would require

• Formal computer system validation (IQ/OQ/PQ) with documented URS, FS, and DS
• Enterprise authentication (SSO/OIDC) in place of the current authentication stub
• Penetration testing and a data protection impact assessment
• Model evaluation against a curated narrative benchmark, with drift monitoring
• Integration with an established safety database rather than standalone storage
• Defined SOPs covering reviewer qualification, escalation, and periodic review

Important disclaimers

• Not a medical device. PV Sentinel is a drafting and workflow assistant, not an autonomous clinical decision-making system.
• Human oversight is required. All AI-generated content must be reviewed and approved by qualified professionals before any regulatory use.
• Validation required. Formal computer-system validation (IQ/OQ/PQ) must be completed before use in a regulated production environment.

Contact

Wayne Kearns ΓÇö wayne.kearns@nortesconsulting.com ┬╖ linkedin.com/in/waynekearns

Happy to walk through the private repository and a live demonstration on request.

---

_{PV Sentinel ΓÇö because the fastest safety narrative is worthless if it loses the patient.}