Insights
Field notes on shipping AI inside real systems.
Latest
Fraud detection on a public dataset: the metric that matters when 0.17% of rows are fraud
We trained four classifiers on a public card-transaction dataset to learn one thing: with fraud this rare, the number everyone quotes is the wrong one. What we measured, and what a production system would still need.
Read post
Your first AI project: pick the win you can measure
The first AI project decides whether there is a second one. Start behind the firewall, on data you already have, with a human gate — and pick something you can measure in weeks.
We rebuilt our own hiring pipeline with AI
Two stages of our own recruiting operation depended on scarce specialist time: assessing senior engineers, and turning interviews into client-ready profiles. We redesigned both around AI. What we built, what it took to trust it, and the numbers.
Documentation that keeps itself current: the agent proposes, your team approves
Docs drift because updating them is nobody's job. Give the job to an agent that detects the drift, drafts the fix, and opens a pull request — and keep the approval where it belongs.
Computer vision on the cameras you already have: what two proofs of concept taught us
Counting people and vehicles, tracking occupancy, checking for a hard hat — the models are ready and the cameras are already installed. The work is placement, thresholds, drift, and deciding what never gets recorded.
Support ticket triage with AI: classify, enrich, route — and measure the reassignments
The first minute of every ticket is spent deciding what it is and who should see it. That minute is automatable — and the number that tells you if it worked isn't speed, it's how many tickets bounce.
An internal knowledge assistant is only as good as its last wrong answer
People ask in Slack because searching is slower and less trusted. A knowledge assistant earns its place the same way a colleague does: by being right, saying "I don't know," and knowing who's allowed to see what.