
I write about AI the way I'd explain it to a smart friend on a Sunday walk.
My read on the seven companies pacing the AI economy. The loudest numbers measure money pouring in, not value coming out — so I separate audited profit from run-rate, trace where capital really pools, and name the game each one is playing. The moat of 2026 isn't a smarter model; it's the line from a dollar spent to work shipped.
When engineering says "we need a validation layer," design says "users need to feel in control," and business asks "what's the ROI at 10x scale" — I make each feel understood AND challenged, then synthesize a path that serves all three.
Shipped enterprise AI assistants with OpenAI APIs, multi-database RAG, and hallucination guardrails. Built eval frameworks for response accuracy in safety-critical industrial operations. Designed context engineering pipelines that survive production.
See the workCoordinated 30+ cross-functional stakeholders across Android, iOS, head-mounted tablet, and Web. Designed progressive trust patterns with a human-centred lens because plant operators don't trust a system when their life is at stake.
See the workPotential $100M+ opportunity. Launched first Alpha 0-1 MES for battery gigafactories. Led consulting motions & pre-sales for product engineering services with 85% win ratio in the past.
See the work12+ years shipping enterprise AI at Fortune 100 scale. From 0-to-1 MES for battery gigafactories to Gen AI assistants for plant operators.
The career narratives, the products, and the journey.
118 deep dives across Evals, Agentic Systems, Harness Engineering, and the AI PM OS.
Written the night a pilot broke, a metric lied, or an assumption snapped — before the lesson got polished into a slide.
One idea at a time. Slow enough to sit with.
The 5% that compounds, without the 95% that doesn't.
The model sets the ceiling. The harness sets the floor.
Every agent failure you've ever investigated had the same autopsy report. Root cause: the model. Actual cause: something else entirely.
The demo passes. Production fails on the same prompt. Teams spend a quarter chasing a better model. A week-two hire rewrites the retry, tightens the JSON, and cuts tools from thirty-one to nine. Errors drop by half in five days, on the same model they started with.
That's the harness. The scaffolding around the model that turns its reasoning into reliable action. Three of the four failure signatures point at it. None of them are the model's fault.
Read the Full Note