Twelve years of putting machine learning where being wrong is expensive — healthcare claims, industrial drawings, enterprise knowledge. Today I design agentic architectures that read a specification, decide what to do about it, and carry the work through to something a reviewer can sign.
Three system archetypes I've shipped repeatedly, and what they're actually worth once they're live.
Software that owns a task end to end. It reads the goal, chooses its own tools, recovers from a bad step, and hands back work a specialist can approve — rather than a draft somebody has to redo by hand.
Answers an auditor would accept. Retrieval grounded in the organisation's own material, with citations back to source, honest refusals, and guardrails that stop a confident wrong answer before it reaches a person.
Models that survive contact with real traffic. Predictive and vision pipelines built to a latency budget, watched for drift, and retrainable by the team that inherits them without me in the room.
Systems, not job titles.
An LLM system that reads healthcare measure specifications and writes the validation scenarios a QA team would have spent weeks drafting.
A predictive analytics system for CMS Star Ratings — treating each quality measure as its own forecasting problem instead of one blended score, so a plan can see which measures are moving before CMS reports them.
A computer vision pipeline that turns industrial P&ID drawings into structured, queryable engineering data.
Longer essays on artificial intelligence, learning, and purpose.
A plain-language tour of what artificial intelligence actually is — and the ANI → AGI → ASI ladder that maps the road ahead for the human species.
Read 4 September 2025Why Elon Musk calls the conventional system boring and unnecessary, and what an activity-driven, problem-first education could look like instead.
Read 21 January 2026Your primary purpose is to make other lives better. Your secondary purpose is the thing you get good at to do it — six ways to find yours.
ReadIf you're building something that has to reason and not just score, I'd like to hear about it.