Four case studies, same pattern: AI work that reduces operational drag, makes quality measurable, and creates commercial leverage.
Each follows the same shape — the problem, the key product decision, the measurable result, the business implication. Built AI-native. Held to one design principle: simple on the surface, rigorous underneath.
Replaced panel recruitment for brand research — faster, cheaper, and still statistically defensible.
Panel recruitment is the bottleneck of every brand tracking study — weeks to field, dollars per response. I built a multi-agent pipeline to generate synthetic respondents, and — because a plausible chatbot with statistics pretensions is worse than no data — a statistical-validation layer that reproduced 100% of the strong-significance drivers (p<0.01) found in real panels and 65–75% at the p<0.05 level in multivariate analysis.
The validation is the product; the generator is the commodity. Anchored an enterprise commercial engagement as lead technical.
An AI tool that replaced a specialist-gated 3-week workflow with a 3-day one.
Survey design was a specialist bottleneck — the problem wasn't capacity, it was expertise. I built a tool that constrains the intake, not just the output: a wizard that guides non-specialists through research-methodology guardrails before any LLM runs.
The key technical choice was keeping the methodology rules outside the prompt as a validator, not inside it as prose. Adopted by the team, presented firm-wide as the company's first AI transformation initiative.
A year of product-and-business-model discovery — reshaping a cross-border procurement problem from marketplace into concierge.
Small Australian builders overpay for materials because direct-sourcing carries trust, compliance, and coordination risk they can't afford to manage. I spent a year finding out what they'd actually pay for — twelve documented pivots, persona interviews, an AI extraction layer that turned construction drawings into structured quotes, and RAG validation against Australian Standards.
The biggest decision wasn't the tech. It was the pivot from marketplace to service-led concierge — three months of code killed, a validated value proposition in its place. Now piloting with two local builders, converting per-transaction through consulting.
An open-source multi-agent pipeline that reads the job market daily, honestly, and at the cost of a coffee.
A five-stage pipeline — embedding similarity, cross-encoder rerank, three-agent scoring panel, tailored material generation — where cheap filters run first, expensive reasoning runs last, and a deliberately adversarial Devil's-Advocate agent keeps the scoring honest.
95% of candidates are filtered at zero API cost. The feedback loop closes it. Open-sourced as a reference architecture for cost-aware, eval-honest LLM systems.
I'm Harry — an AI-native product manager with data science rigour. Data science gave me the rigour on outputs; delivery management gave me the shipping discipline; current AI product work puts them together. I use AI tools daily to stretch what one person can ship.