The work
Start with the current ventures, then wander through SThree, Datamise, Oxford, Astroscale, and the experiments that shaped how I think about AI systems.
The future-facing company story: patent intelligence and dental voice operations.
Current founder venture
AI-assisted pharmaceutical patent-intelligence support where the AI is kept in check by fixed rule checks and human review.
This is Peter in builder mode: text, drawings, rules, and review all working together in a domain where guessing is not an option.
Pre-launch founder venture; not legal advice, a legal opinion, public availability, customer adoption, or revenue.
Current founder venture
AI voice operations for dental practices, designed around missed-call recovery, booking support, graceful failure, privacy, and launch gates.
A hands-on voice product: live calls, awkward edge cases, graceful failure, operator visibility, and the discipline to launch carefully.
Pre-launch founder venture; not public availability, customer adoption, revenue, clinical safety, or receptionist replacement.
Aurum at SThree: coordinating teams of AI agents, scoring systems that learn to score better over time, and keeping it all reviewable and controllable.
AI at SThree
Major employer work at SThree: teams of AI agents coordinated together, realistic test candidates, scoring systems that learn to score better over time, reviewable steps, bias checks, deliberate stress-testing, and production control.
Shows Peter at his most systems-minded: ambitious agentic AI with review, control, and stress-testing built in from the start.
SThree employer work with employer-IP boundaries; not an independent venture, public implementation guide, legal-compliance certification, or real-applicant validation claim.
Datamise and SettleMise show the client work behind the current ventures.
Datamise client work
Datamise client work: owned claims infrastructure and SettleMise, a reviewable equal-pay valuation engine.
The grounded, human-scale build story: a union operation, messy real-world data, and a calculation pipeline checked to the penny across 50 runs.
Supports human legal and accounting review; does not decide liability, replace professional judgment, or claim legal outcomes.
Oxford and Astroscale show the hard-domain technical foundation beneath the AI product work.
Deep technical authority
Medical AI research that reads full digitised microscope slides: finding individual cells, sorting them into eight types, mapping tissue regions, and spotting patterns across a 3-million-sample dataset.
Shows the research habit Peter brings into product work: measure carefully, keep the biology in view, and say what did not work too.
Academic research in mouse placental histology; not a clinical diagnostic, clinical deployment, or patient-product claim.
Hard-domain computer vision
Past employer work in computer vision that pins down a satellite’s position and orientation in space, and AstroGAN, which makes cheap simulated images look like real satellite photos so the systems train on better data.
Shows deep technical range in physical-world AI where domain shift, geometry, and measurement matter.
Past employer work; no claim that it flew in space, that the simulation-to-real trick works everywhere, or that the full system was validated end to end.
The craft of knowing when to take large language models out of the live system and state the trade-offs openly.
Production craft
SThree’s system for tagging skills on a CV, rebuilt to look up close matches, apply fixed rules, and use a tiny fast classifier — so it no longer calls a large language model every time it runs.
Shows production judgment: knowing when to take the large language model out of the live system, and naming the quality trade-offs instead of hiding them.
Supporting SThree work; not the headline project, and not a claim that a tiny model can replace large language models everywhere.
Private product R&D, benchmarking, evaluator research, and exploration work that show how Peter checks AI systems before trusting them.
Private product R&D
A private multi-agent assistant in daily use for capture, durable memory, proactive review, action, and quality monitoring.
This is Peter building for his own life first: memory, proactive loops, product taste, and a quality system without exposing private context.
Private single-user dogfooding; not employer work, customer adoption, public SaaS, or external certification.
Evaluation depth
Independent benchmark work comparing different ways of running AI research agents — checking that the automated scorers can be trusted, verifying every cited source, and testing fairly for genuine differences.
Shows method discipline: building fair comparison setups, reporting the results that did not work, and resisting leaderboard showmanship.
Independent research artifact; not employer work, accepted publication, or a claim that one architecture wins universally.