PharmaTools.AI – Applied AI for Pharma & Health
Hi, I’m Nick Lamb

I build AI for domains where being wrong isn’t an option.

Twenty years writing regulated medical communications taught me what “verified” has to mean. Now I ship AI systems that meet that bar – including an MHRA-registered Class I medical device and an open-source evaluation framework used beyond healthcare.

Applied AI engineer Evals & verification Regulated domains
See the evidence View résumé
Nick Lamb
Nick Lamb, PhD, CMPPFounder · 20+ years in medical communications
approach.md
// My approach
1 Ground every claim in evidence
2 Verify deterministically, not by vibes
3 Protect privacy by design
4 Ship. Measure. Iterate.

// Outcome
✓ AI that earns trust in high-stakes domains
✓ VerifiableMHRA Class IRegistered medical device (Patiently AI) ✓ Verifiable3,000+ / monthDownloads across the open-source suite (npm · PyPI) ✓ Verifiable20+ yearsRegulated medical communications ✓ VerifiableMCP Directory ×2Redacta & PubCrawl – official Anthropic directory

01Evaluation first

Measured, not claimed.

Anyone can demo an AI system. The hard part is proving how it behaves – deterministically, adversarially, and at a cost you can defend. This is the work I care most about.

OpenGATE

1.9K downloads/mo

Open-source verification framework for evidence-grounded AI. Deterministic grounding and extraction checks with no LLM judge – reproducible, auditable, and free to run on every answer or every commit.

“Don’t trust AI. Verify it.”

npmPyPIMCP serverDockerGitHub Action
Explore OpenGATE →Read the paper →

Groundwork

GitHub ↗

Deployment-readiness harnesses for document QA and structured extraction – built for the small team with a working prototype, real users waiting, and no ML engineer.

“From AI prototype to production.”

2 archetypesnpmMCP serverBuilt on OpenGATE & Redacta
Explore Groundwork →

Redacta Gauntlet

GitHub ↗

Adversarial evaluation harness for Redacta’s PII engine – hostile formats, near-misses, prompt injection and context leakage, with a published threat model. Because a single missed identifier is a privacy breach.

28 adversarial casesPrompt injectionRedacta DOI 10.5281/zenodo.21115605
See the Gauntlet →
refcheckr – model eval run
ModelExactHalluc.$ / 1K
sonarprevious production52.5%5.8%$6.75
sonar-proproduction now61.7%✓ 2.4%$14.29
sonar-reasoning-pro82.8%15.1%~$11.30
Verdict accuracy (exact, six-point scale) · passage hallucination · measured cost per 1,000 claims. 60 verdict pairs × 3 repeats, 2–3 full runs per tier, averaged. The decisive axis is hallucination – invented evidence is the worst failure for a verification tool, so production runs sonar-pro despite the higher cost. Full methodology, charts and raw scorecards →
5.8% → 2.4% Passage hallucination halved by choosing the production model on measured evidence, not reputation
WhitepaperDesigning Trustworthy AIWhy high-stakes systems are converging on verification-first architecture. Paper · ZenodoEvaluation as InfrastructureDeterministic regression gating for evidence-grounded AI, across four production systems. Article · dev.toRAGAS vs DeepEval: same fabricated answer, 0.0 vs 1.0A controlled comparison of faithfulness metrics against a deterministic baseline. PlaybookMedical Writing AI PlaybookWhere AI fits – and where it doesn’t – inside pharma compliance codes.

02Things I’ve built

One thesis, shipped five ways.

Every product answers the same question: how do you make AI dependable enough for healthcare?

All projects on GitHub →
→

Patiently AI

MHRA Class I · Web, iOS & Android · 18 languages

Translates clinical notes into clear, actionable explanations for patients – built to regulatory standards as a registered medical device.

 →

RefCheckr

Verification · Web · Word add-in

Automated claim and reference verification for medical and scientific content, with its model choices benchmarked in the open.

 →

Redacta

Privacy · On-device · Self-hosted · Anthropic MCP Directory

Pseudonymises patient identifiers before text reaches an AI model – reversible, NHS-aware, benchmarked under adversarial attack, and deployable into your own Kubernetes cluster.

 →

PubCrawl

Retrieval · Anthropic MCP Directory

Gives AI assistants direct access to PubMed, Europe PMC, clinical trials and drug labels – real papers, no hallucinated citations. A one-click remote connector: no install, works across Claude on web, mobile and desktop.

 →

MedCheckr®

Compliance · 11 pharma codes

Checks promotional claims against 11 national and international pharma codes of practice in seconds, with citation-grounded regulatory feedback.

More experiments →

PosterLens, HushMap, BiomarkerFinder, StudyDiff, TrialGen and other prototypes, demos and detours live in the Labs.

Research · Observer Zero

150 simulated universes, populated by AI scientists. When the laws of physics secretly changed, they noticed – and never once believed it.

An open research programme on how AI agents do science when the ground truth is hidden from them. Published study, open dataset, runnable platform.

Explore the research programme →
0 / 40societies concluded that a law of their world had changed

04Recognition

Independently acknowledged.

Communiqué 2025Winner – Progress in Healthcare & Scientific Communication
MHRARegistered Class I medical device
PMEA 2025Double win – Patient Education & Innovation
Open TargetsHackathon Audience Choice 2025
Frontiers in Digital HealthValidation study in peer review
7 startup programsGoogle · AWS · Microsoft · Anthropic · NVIDIA · Perplexity · ElevenLabs

05Consultancy

Applied AI consulting

Adopt AI with clarity – not hype

Hands-on advisory for pharma, medtech, and healthcare teams – grounded in building and shipping AI products, including an MHRA-registered patient-facing medical device. From focused prototypes to patient-facing AI strategy.

Learn more →
Patient-facing AI advisory
Evals & verification design
Compliance-aware AI design
RAG & knowledge workflows
API & product integration

Interested in working together?

Explore how applied AI can support your team – from compliance to patient communication.

Get in Touch Explore AI Consulting