Skip to content
Welzin
Talk to us

Insights

From insight to impact.

Practical, opinionated writing on building AI and data systems that hold up in production.

Showing 15 of 34

GenAI & Agents

Agents that do the work, not just chat

A 95%-per-step agent completes a 10-step workflow 60% of the time - that math, not model quality, is why agents fail. Where agents genuinely work in 2026, and the verification-first architecture that gets them there.

9 min read

AI Strategy

Own the metric, not the model

Model metrics are proxies, proxies degrade under optimization (Goodhart), and the fix is organizational: metric trees, guardrail pairs, attribution via experiment, and a named owner for the business number.

10 min read

AI Strategy

Buy, build, or fine-tune

Prompt, then RAG, then fine-tune, then distill - and buy the commodity while building only the differentiation. The decision framework, the real break-even numbers, and the 3-5x TCO trap.

9 min read

Engineering

The quiet cost of the prototype that never ships

MIT found 95% of GenAI pilots deliver no P&L impact; 4 of 33 POCs reach production. The compounding costs of pilot purgatory - trust, talent, pilot fatigue - and how the shipping 5% sequence differently.

9 min read

Data & Analytics

The feature that quietly broke your model

DoorDash found feature mismatches of 35.7%; Google Play gained 2% installs by fixing one skewed feature. Why models get blamed for pipeline crimes, and the input-first observability that catches the real culprit.

9 min read