Track everything, silently
Bolt a monitoring agent onto every IDE and you get compliance-shaped data: engineers route around it, sanitize their prompts, or quietly stop using AI where it's watched. You end up measuring fear, not skill.
Plaiback Kata — practice, not a performance review
Run it again next quarter. Watch the numbers move.
Kata isn't a one-time audit — it's a repeatable exercise your engineers run each quarter, so improvement shows up as a trend, not a guess. Anonymized by default: you see the spread and the trend, never a dossier on one person.
Scoreboard above is illustrative — each row compares against that participant's own last run.
The problem
Nobody wants to be watched. Nobody wants to fly blind either.
AI usage inside an engineering org is either invisible or surveilled — and both leave you guessing about where your AI budget is actually going.
Don't measure at all
You're paying for Copilot, Cursor, and Claude seats across the whole org with no idea which teams are getting real leverage from them and which are burning tokens on trial and error.
There's a third option: an anonymized, opt-in practice kata, run quarterly so the goal is visible improvement, not a permanent record. Cohort-level insight instead of a dossier on each engineer — identity revealed only when someone chooses to be.
See how it worksFrom cohort to coaching, quarter after quarter
Same replay-driven engine as hiring, run on a cadence so you see improvement, not just a single snapshot.
Run a kata each quarter
Pick a standardized task (or bring your own), invite a cohort — a team, a business unit, or the whole company — and set the budget and window.
Anonymized scoreboard
Everyone is graded and judged the same way hiring candidates are — but results default to anonymous. The spread and the patterns are visible; identities aren't, unless someone opts in.
See what actually drove it
A coaching report compares top and bottom results — prompting discipline, model choice, context handling — and turns it into concrete advice, not a leaderboard to hide from.
Run it again — watch the trend
Each quarter's kata lands next to the last one, so you can see whether prompting discipline, model choice, and cost per outcome are actually improving — not just where things stand today.
A practice session, not a performance review
Same replay-driven signal as hiring, pointed at your own team's day-to-day AI workflow.
Anonymized by default
Results are pooled and ranked without names. Reveal is opt-in — top performers can choose to be credited; nobody is exposed for scoring low.
Improvement you can see
Kata runs on a cadence, so this quarter's result sits next to last quarter's. The point is the trend, not a single snapshot — and it shows whether the coaching is actually landing.
Same engine as hiring
The event log and LLM judge are the same ones already proven on candidate assessments — not a separate, unproven tracking tool bolted on afterward.
Cost & model-allocation signal
See who's escalating to expensive models out of habit versus need, and who's getting more done on cheaper ones — the real driver behind your AI spend.
Replay showcase
Opt-in top runs become a growing internal library your team can learn from — a compounding asset, not a one-off report.
Your stack or ours
Use the standard kata catalog, or author tasks against your own codebase and patterns.
Run your first kata
We're onboarding teams by hand while we learn what works. Join the wait list and tell us the size of your org — we'll get a cohort set up.