Building systems that can recognize when they’re wrong

I built and led healthcare products at Doximity, Transcarent, and CancerCompass, and co-founded Andwise. Now I build tools for the questions that come after launch: has the evidence behind a decision changed, what was an AI agent approved to do, and did a correction reach the people who acted on the error?

Portrait of Kanav Jain

Healthcare systems  · Human–AI interaction  · Institutional accountability

300,000+

calls on an average workday

9 years
after I left, still in daily use
Doximity Dialer, which I originated as founding product lead. Doximity’s report for October–December 2025. Check or challenge this figure →

At Epic, the software was live, and a regulatory change or a critical bug still had no defined route to the person who could fix it, so I wrote the escalation path. I now study that gap in other institutions: when a tool they rely on gets something wrong, can the people affected challenge the decision and get it corrected?

Research →

For sale

NextConsensus

Estimates how likely a guideline, label, or coverage decision is to change by a set date.

NextConsensus share card for FF-011: “Will FDA approve ivonescimab for EGFR-mutated non-small cell lung cancer after EGFR TKI therapy by Nov 14?” with a PDUFA goal date of 14 November 2026 and a registered probability of 15%
Enlarge image
Registered 2026-09-26 at 15%, compared with a 70% estimate for similar FDA approvals by their deadline. The comparison is an estimate, not a count of past decisions. Unresolved at the 2026-11-10 registry snapshot; the registered rule reads the FDA record no earlier than 2026-11-28.

Experimental

Ambit

Maps what an agent setup can already do, what one more step would unlock, and which of those tools actually work.

Try the capability map

A terminal running “ambit approve” on an agent’s proposal: approved by Jordan, applicable true, with the note that the approval artifact is signed and expires
Enlarge image
The agent drafted this change through Ambit’s tools using Model Context Protocol (MCP), which connects agents to tools. Those tools let it propose a change; approval and application require a person in a terminal. The approval records that person’s declared name, is tamper-evident, and expires after 24 hours.
  • Live

    Refract Replays source revisions to catch a change before the citations that rely on it go stale.

Current work: code and checks →

  1. Essays

    I write at The Crumple Zone about who absorbs an institution’s failures and who can make it correct them. Selected writing →

  2. Proposed AI standards

    I wrote the Ethotechnics framework, an open set of proposed standards for recording approval, challenges, and corrections. ethotechnics.org

Whether a record can carry a real dispute end to end: an agent acts, a person challenges the action, and the correction reaches every decision that depended on it.

This site is the smallest version of that test. Tell me a figure here is wrong and the challenge is listed whether or not I agree; if you are right, every page that printed the figure says so. Corrections →

Anything else: kanav@kanav.net

Revised