Notes from the seam between AI models and the workflows they get bolted onto. Mostly manufacturing and healthcare. Mostly what breaks in the second month, after the demo is over.
In my last post, I explored the critical role of context management in getting the best results from AI. Today, I want to share two real-world examples that highlight a simple but powerful technique you can start using immediately.
Context Engineering: The $100K Skill You
In my last article, “What if GenAI could give MedTech professionals back tens of thousands of hours a year?”, I shared how my experience programming with coding agents. In this follow up, I wanted to share how I approached writing an advanced RAG solution to help MedTech professionals to efficiently
That’s not a hypothetical—it’s a real possibility I’ve been exploring as I build GenAI-powered solutions for regulatory documentation in MedTech and BioTech. MedTech and BioTech are governed by regulators like the FDA, EMA, PMDA, ANVISA, and IMDRF.
I’ve taken up coding as a hobby recently. The last time I wrote code was before C#, GitHub, containers and development pipelines.
Better agent control using hierarchical agent.md/claude.md file structure:
DSPy's BootstrapFewShot optimizer can learn from user feedback to improve prompts automatically - no manual prompt engineering needed.
MedTech advisor in Chicago. Most of my work sits at the seam between AI models and the workflows they were bolted onto — manufacturing operations, regulatory and compliance, supply chain, and the long tail of things that quietly break in the second month after a pilot ships.
The coding here is so I understand the boring half of generative AI well enough to advise on it. I work on agentic systems, eval harnesses, and the integration plumbing that decides whether a pilot survives past the demo. When the model is the easy part, I'm interested in the rest.
When I'm not coding or advising, I'm in the kitchen, in the workshop, or behind a camera lens.