irvin

1–5%

of recorded sessions a supervisor can realistically review today

~3 min

of flagged moments to review instead of a full hour

HIPAA

regulated, with PHI handled end to end

Clinical supervision runs on a time deficit. A supervisor can realistically watch only a small fraction of their supervisees' recorded sessions, so supervision meetings drift toward whatever the supervisee remembers happening. Safety concerns surface late, and the competency documentation that accreditation requires becomes a manual slog. The wedge Irvin bets on is coverage: review the flagged minutes of nearly every session instead of spot-checking a few.

Great Oak brought the clinical depth and real supervisors to learn from; I turned that into product decisions. The evaluation framework is Bernard's Discrimination Model, and the most unusual design artifact was the AI's voice itself. We specified Irvin like a seasoned, trusted supervisor: encouraging but direct, always framed as attempt, gap, alternative, question. It has a hard cap on how many Socratic questions it may ask and an explicit forbidden-language list, because in a clinical tool the tone is part of the safety model.

The recurring decision was how much AI judgment to put on screen, and the answer kept being less. A radar chart of competencies looked impressive and read terribly, so it became a plain top-three and bottom-three skill list. A Draft and Submit button pair collapsed into one state-aware button with a status pill. The AI's job here is to point a supervisor at the right three minutes, not to perform intelligence. Less AI judgment on screen, more UX.

The hero flow is session review: the transcript on one side, Irvin's analysis on the other, with an overall competency score, per-exchange analysis, a speaker-activity timeline, and Critical Concerns flags that surface risk right away, because some things shouldn't wait for the next supervision meeting. Supervisors read each session two ways, skill by skill and then as a whole, so feedback covers both the craft and the hour the counselor actually ran. Around it sit the supervisee flow, from upload through draft to a locked submission with licensure-hours logging, and the supervisor flow, with a Needs Attention queue and per-student competency trends.

Irvin went live in production in April 2026: frontend on Vercel, backend and database on AWS, and every AI inference routed through AWS Bedrock under a signed BAA, so protected health information never leaves a HIPAA-covered environment. Bridger is still on the product, and it keeps shipping. Early supervisors describe reviewing more sessions in a week than they previously managed in a month.

Healthcare set a different bar: trust over delight. Every meaningful decision went into a written log with its reasoning, which sounds bureaucratic and turned out to be freeing, because nothing had to be re-argued twice. And the brand pulled its weight: calm greens, Crimson Text, real photography of counselors at work, and language that mentors rather than monitors. Irvin feels closer to a therapist's office than a dashboard, which is exactly where its users live.