I'm especially interested in the layer around the model: how software tracks what changed, preserves provenance, knows when assumptions are stale, and identifies repeated human work that should become durable system capability.

• Ambit — models what an agent system can actually do across models, tools, machines, permissions, and humans—and where it still gets stuck.
• Refract — tracks changes in source material with provenance, so downstream systems know what changed rather than repeatedly rereading the world.
• NextConsensus — forecasts specific actions by healthcare institutions from accumulating public evidence.
• whether.work — turns macro and capital conditions into a weekly answer for startup leaders: how aggressively to hire, spend, raise, and expand, with explicit stop and reopen conditions.
• Stims — programmable audio-reactive visuals, partly exploring how software can give users more control over sensory environments.

A lot of AI capability is limited less by the model than by the surrounding system: missing context, stale state, weak feedback loops, brittle permissions, and humans repeatedly filling gaps by hand. I'm interested in building that surrounding layer better.

U.S. healthcare has spent 15 years moving decisions into software. My work is about making sure responsibility moves with them. I spent 14 years as a product lead across Epic, Doximity, CancerCompass, Transcarent, and Andwise — building products where decisions have to be traceable, challengeable, and reversible.

• Doximity Dialer — founding product lead. Patients don't answer calls from unknown numbers — that single observation drove the product. Dialer now powers 300,000+ daily visits across 200+ health systems.
• CancerCompass / CTCA Marketplace — led digital products for an oncology navigation platform serving 30MM annual visitors. Cut bounce rate 25%, lifted chat conversions 267%.
• Transcarent — directed care-navigation product across Surgery, Urgent Care, Behavioral Health, and Oncology Care.
• Epic — configured EHR workflows, where a single misrouted alert could bury a critical lab result. The workflow around the clinician was usually the constraint, not the clinician.
• Andwise — co-founded a physician financial-wellness company. Raised $240K, grew to 1,200+ physician users and a 700-member community.
• Georgia Tech — studied RNA folding dynamics, co-authoring a chapter in the ACS Symposium Series at 19.

• refract-org/refract — structured change events from public knowledge sources, starting with MediaWiki revision history
• zz-plant/stims — live at toil.fyi; 1,868 presets, AI generation, CodeMirror editor, mic/tab input
• zz-plant/whether — deterministic engine that turns macro signals into a weekly Boldness Budget and bounded operating rules for hiring, spend, fundraising, and expansion
• zz-plant/ambit — capability graph for agent systems; what they can do and what they may do
• zz-plant/blog — custom publication and newsletter stack
• zz-plant/bread — archive for BREAD, a music party I co-founded in San Francisco

If you're working on a healthcare AI product that needs to survive real workflow, review, audit, override, and mistakes, I'd like to hear about it.

• kanav@kanav.net
• kanav.net
• github.com/zz-plant