Get in touch

How do we make things better without making them harder to change?

I bring new AI capabilities through hospital review into products still in use after I leave. What I build now asks a second question: after a system works, can the people who depend on it still see how it decides, object when it is wrong, and change it?

See current work View résumé → Healthcare case studies →
Portrait of Kanav Jain

AI product leader, engineer, and researcher · Chicago

calls on an average workday
300,000+
since I left, still running
9 years

Doximity Dialer, which I originated as founding product lead. Doximity’s own figures, February 2026.

Selected work

Doximity

Decided by approval 2013–17

The patient sees the office, not the cell phone

Adoption may come before official approval only while the risk stays with the person choosing. When the risk is clinical, approval comes first.

  1. The situation Physicians called patients from personal phones, giving the number away for good or showing “Unknown Caller.”
  2. My choice I proved caller ID over a weekend and launched on a vendor that did not require proof of number control.
  3. What followed Hospital authorization followed adoption, and the route was unchanged when I left.
Read full case study →
Doximity Dialer “Call from” selector set to OFFICE, with a tooltip explaining the recipient sees this number as caller ID
The identity decision in the product’s own words: the clinician picks which number the patient sees, with the office line preselected.
CancerCompass / CTCA

Decided by attention 2019–21

Scaling oncology navigation for 30 million annual visitors

A frightened person is not short of information. The scarce resource is attention and trust, so the product’s job is the next step and a person to call.

  1. The situation A family facing a new cancer diagnosis needed the next step and a number to call. The site gave them encyclopedia entries.
  2. My choice I relaunched around plain-language care pathways and nursing triage, with product and nursing reviewing real questions together.
  3. What followed Chat conversions rose 267% and bounce rate fell 25% across 30 million annual visitors.
Read full case study →
Transcarent

Decided by accountability 2021–22

Four specialty programs on one routing architecture

One system can carry many programs once every decision point has a named owner.

  1. The situation High-complexity cases in surgery, oncology, and behavioral health were stalling in exception queues.
  2. My choice I built one shared routing system instead of four custom ones, and stated the trade: slower to build, faster to scale.
  3. What followed Four programs launched on shared infrastructure, and routing delays stopped being the primary bottleneck.
Read full case study →
See all case studies →

Current work

The record above is why these exist. Each project takes something people rely on and keeps it correctable: an estimate that resolves on a date, a permission that can be withdrawn, a source change that reaches the claims citing it.

Commercially offered · public forecast registry and review methodology; no resolved forecasts or paid engagements yet

NextConsensus

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

An illustrative NextConsensus claim record: at revision 4 the word “contested” is struck from “clinical benefit remains contested,” above the claim’s emerged, contested, and authority dates
Enlarge image
Illustrative record. At revision 4 an editor deletes “contested,” and a disputed claim reads as settled. NextConsensus gives teams dated records like this one.

Experimental · open source, with a live demo and a roadmap that says what is not built

Ambit

Distinguishes the tools an agent can reach from the actions it is authorized to take.

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 over MCP, which has no tool to approve or apply it. A person approves it in a terminal under a declared name, and the approval is a tamper-evident record that expires after 24 hours.

  • Live · open source, with a scheduled observation bot and a 15,926-event public test set

    Refract Records when a source changed and what changed.

Current work: code and checks →

Writing and standards

The standards and many of the essays ask the same question: would a system still look like it works if we stopped counting the people covering for it?

  1. Research and essays

    From a first-author ACS Symposium Series chapter on RNA thermodynamics to 230+ essays at The Crumple Zone on how institutions distribute burden, absorb failure, and decide who can exist inside them. Selected writing →

  2. Proposed AI standards

    I wrote Ethotechnics’ framework for recording approval, challenges, and corrections. ethotechnics.org

What I want to test next

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. NextConsensus, Refract, and Ambit are three parts of that test, and the record format is published as an open standard so others can check it.

One essay that makes it concrete People Are Not the Error-Correcting Layer How informal human repair hides a software defect from the people who could fix it.

If you run a system that depends on people covering for it, or you are building the record that would stop it, I would like to hear what you are finding. kanav@kanav.net

Revised