AI Product Manager — Agentic & LLM Systems · 15 years in Product/SaaS, 8 years as Product Owner / Product Manager · Sovereign & Regulated AI

An open-source-intelligence platform built for a defence innovation challenge: it ingests public signal, scores its credibility and produces cited analyst briefs, entirely inside an air-gapped, CPU-only enclave. I owned it end to end as product owner and builder: problem framing, architecture, build-versus-buy, evaluation design and delivery. It proved that a grounded system can be held to an evaluation floor in code, citing every claim it makes and saying nothing when the corpus cannot support an answer. Full case study on evaluation close.

An internal enterprise AI platform at a global asset manager, giving teams a governed route to large-language-model capability, retrieval over internal knowledge and agentic workflows inside the firm's regulatory boundary. I was the product owner: discovery, backlog, prioritisation and the stakeholder path from platform teams to the business users adopting it. It proved that an LLM platform can be product-managed like any other platform, with adoption, cost-to-serve and guardrails as first-class requirements rather than afterthoughts. Full case study on employer clearance.

A conversational assistant on the personal-investing platform of a global asset manager, answering general investing and account-servicing questions for retail customers in a regulated market. I worked on the product side of its delivery, with the platform above as its foundation, alongside engineering, compliance and the customer-service teams. It proved that a retail-facing LLM assistant can ship in regulated finance with explicit boundaries between information and advice, human oversight and stated limits in the first message. Full case study on employer clearance.

A proof-of-concept operations engine for a direct-to-consumer brand on a hosted commerce platform: live store integrations, order and cash-on-delivery reconciliation, and content tasks routed to different model tiers by the work each task needs. I owned the proof of concept from problem statement to success criteria, and built it with an AI-assisted delivery method. It proved that routing by task, with the most capable model reserved for build and judgement work and cheaper models for bulk and tagging, meets the acceptance bar at a fraction of a single-model cost. Full case study on client consent.

A ticketing and travel platform for a festival organiser, which began as a boarding management system and grew into identity and role-based access, a catalogue, checkout, QR tickets and an operations console. I owned product discovery and delivery: the product requirements document, the phase plan and the sign-off gate each phase had to pass. It proved a disciplined discovery-to-delivery method on a product with no AI in it, which is the point: the method is the same whether or not a model is involved. Full case study on founder consent.

A quiz game that tests AI product-management concepts, built from a private syllabus of thirty-plus concepts covering agent patterns, evaluation, memory, guardrails and delivery. I designed the question bank's structure and the game modes, and built it with an AI-assisted delivery method. It proves that a learning artefact can be turned into something a reader can use in two minutes. Live tool once hosted.

A retrieval layer over a private technical library, exposed to a coding agent as a tool through the Model Context Protocol, so the agent can cite the library instead of guessing. I owned the design: the extraction pipeline chosen on measurement, the chunking and index, and the tool contract the agent calls. It proves that retrieval-augmented generation is a product decision about what the agent is allowed to know. Private tool; the tooling and method may be described, the corpus never.