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.

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. Everything below is a working piece of that surrounding layer.

Full-screen visuals that move to whatever you're listening to. Every scene is a MilkDrop preset — a small visual program — compiled from raw .milk source in the browser, with no conversion step and no server.

2,679 presets, a live CodeMirror editor with MilkDrop completions and diagnostics, WebGL2 with guarded WebGPU, and audio from YouTube, mic, a file, or a browser tab. Released into the public domain.

Live at toil.fyi · zz-plant/stims · Unlicense

Turns document histories into claim-state timelines. Given a page, Refract emits a structured event stream of what changed, when, and how — every sentence introduced, modified, or removed; every citation that shifted; every revert and edit cluster.

No model and no inference in the observation path — the same source produces the same events every time. A hash-pinned ground-truth corpus of 16,146 events across ten benchmark pages ships as a release asset. There's a Python SDK, a web explorer, and an MCP server so an agent can query the event stream directly.

refract-org/refract · Docs · AGPL-3.0

Models what an agent system can actually do across models, tools, machines, permissions, and humans — and where it still gets stuck. What an assembled agent can do is kept separate from what it may do.

Click a node to inspect its dependencies, verified evidence, and blast radius; simulate an outage to see what stops working. Runs as a meta-MCP server, so agents can query their own capability surface before acting.

zz-plant/ambit · Interactive demo · MIT

Turns macro and capital conditions into a weekly answer for startup leaders: how aggressively to hire, spend, raise, and expand — with the stop and reopen conditions written down before they're needed.

Deterministic: the same inputs produce the same call. Every posture carries an explicit trip condition and a reopen condition, so a reversal is a rule firing rather than a change of mood.

Live at whether.work

A publication about how institutions distribute their own failures. Hospitals, workplaces, platforms and bureaucracies stay calm at the center by pushing the shock outward; the essays follow it to whoever ends up absorbing it.

231 essays so far, on a publication and newsletter stack I built and host myself.

Live at thecrumple.zone

A governance commons built around one question most frameworks skip: can this system be stopped while it's harming someone, by someone other than its owner?

Standards, checklists, worked examples, and diagnostics, organized so you enter with a real situation — a live decision, an incident, a policy gap — and leave with named owners, clocks, and evidence rather than a maturity score. Seven proposed standards, crosswalked to NIST AI RMF, ISO/IEC 42001, and the EU AI Act.

To see the operating model applied, the Reliance Lab walks one composite deployment — a clinical summary assistant, six months in, with new information on the table — through the decision of whether to keep relying on it.

Live at ethotechnics.org · zz-plant/ethotechnics.org

Decision briefs for contested healthcare claims. NextConsensus forecasts specific actions by healthcare institutions from accumulating public evidence — tracking how a claim moves across evidence, labels, payer policy, guidelines, safety signals, and public dispute, then packaging the read as a source-backed review brief.

The goal is not to declare truth. It's to show what changed, why it matters for a pending decision, and where a reviewer still has usable recourse. Refract is the open-source observation engine behind part of this workflow.

nextconsensus.com · github.com/nextconsensus

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. Built the identity layer that let a physician call a patient from their own phone while the patient saw the office number — so the call got answered and the physician's cell number stayed private. Nine years after I left, Dialer carries 300,000+ calls on an average workday across 250+ hospitals and health systems — #1 Best in KLAS Telehealth Video for five consecutive years.
• 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 — owned what happened to eight client organizations after installation: a 477-bed cancellation risk worked back to a reference account, a 134-year-old health system taken online, and the federal quality-reporting escalation path. The workflow around the clinician was usually the constraint, not the clinician.
• Andwise — co-founded a physician financial-wellness company: software that read a physician's employment contract and flagged the clauses worth arguing about, with a named reviewer behind every analysis. Raised $240K, grew to 1,200+ physician users and a 700-member community; wound down in 2024.
• Georgia Tech — studied RNA folding dynamics; first author on a chapter in the ACS Symposium Series (Vol. 1082, 2011) on the thermodynamics of sRNA–mRNA interactions.

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