blog.
How I Run My Job Search in AI Era
Mass applying still works if every application is precise. Here is my step-by-step setup: a detailed profile, a match score gate, an ATS check, and three MCP connections, plus what I got wrong.
Evaluating Agents: Outcome First, Then the Path
An agent's last message is its own report of what happened. How to grade the outcome, diagnose the steps, and why customer-facing agents need pass^k.
Your Users Are Already Grading Your AI
Regenerates, edits and overrides are a free eval running in production. How to read them, and why experts warn against turning them into a score.
Who Checks the LLM Judge?
An LLM judge is a model too. How teams check a judge against one expert's labels, which biases to test for, and a validation plan that fits on a page.
The PM Owns the Error Log
The most useful eval work a PM does is deciding what counts as a failure and turning each real miss into a test. Error analysis, ownership and the anti-patterns.
Most of Your AI Evals Should Be Boring Code
Before using a model to grade a model, write the checks a rule can catch. What the industry says about code graders, LLM judges and human review.
When the Roadmap Meets an LLM: How I Reorder Priorities
AI features carry more uncertainty than typical roadmap items. Here's the framework I actually use to decide what's worth this quarter.
let's build something. ✦
Happy to talk about products, AI and the calls behind them.