What I do, where I've done it, and the work that came out of it.
I write the documents work actually gets built from. Requirements, PRDs, roadmaps, sprint plans, and the architecture calls sitting underneath them. Most of it starts with a founder or partner session that I run myself, and ends as tickets a team can pick up on Monday morning. The pipeline that does the middle of that is written up here.
I came into product through design, so I still do the work. Interfaces for AI products on web and mobile, design systems, brand identity, and workflow diagrams. When something needs to exist quickly rather than perfectly, I build it in Framer or Webflow instead of writing a brief for it. Six years of it is in the playground.
I write my own tooling. Custom Claude skills, chained agent pipelines, prompt and evaluation design, and MCP and API integrations. Most of it exists because the workflow I needed did not, and the fastest way to get it was to build it myself. All of it is public.
A document leaves too much open. So I stand up a working version early with Claude Code and Codex: the real user flows, the real states, and the whole platform structure, running on mock data. Clients and their users can move through it and see how the product will actually operate before anyone commits a sprint to building it. The property engine is the clearest example.
Before anything gets built I do the unglamorous part. Market and competitor research, customer segmentation, prospect evaluation, and technical deep dives to work out whether a partner and an idea are genuinely a fit. The positioning and the copy usually end up being mine too. Here is one where the research changed what we built.
Most of my work is agentic AI applied to regulated or operationally messy industries. Three I can go deep on: healthcare operations, real estate, and freight and logistics. Each has a shipped product or a case study behind it. Also worked in capital markets, construction, edtech, and sales and customer success. The pattern repeats: high stakes, scattered data, and people who do not want another dashboard.
Three of these in full: the situation, what I owned, and the call I had to argue for.
Four chained agent skills taking client documents to Linear tickets, with approval gates where the blast radius is largest.
In daily use across the product team
Thirty-one weighted factors, personalised by buyer profile, with confidence and source on every one.
Live, used by real buyers
Four candidate products, one recommendation, and the reason I argued against building the obvious foundation first.
Live with design partners
Building infrastructure for AI agents.
Patients forget most of what they are told after surgery. A clinician uploads a recovery protocol; the system extracts the timeline, writes a script, clones the clinician's voice, and delivers it as a narrated avatar video with check-in questions. I defined the generation flow and wrote the spec.
Autonomous analyst agents that pull earnings filings, monitor news, and produce analyst-grade coverage, alongside a matching layer for private share demand. I ran the founder sessions and shaped the use cases into a scoped requirements set.
Every company has a process on paper and a different one in practice. This engine reconstructs how work actually flows from system logs and unstructured artifacts, then shows the bottlenecks and the rework. I built the framework spec and its first vertical module, for freight.
Public funding goes unclaimed because chasing it is a full-time job nobody staffs. The engine filters a large grant corpus on hard eligibility rules, then ranks and explains what's left with draft proposals attached. I owned the roadmap and phased the matching pipeline.
Turbocharged GenAI fine-tuning & inference platform.
