Every one of these projects deals with the same failure: a record that has gone out of date and is still being relied on. A guideline that no longer matches the evidence. A permission that no longer matches what the system can reach. A correction that never reached the people who acted on the error.

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, built to keep authority, evidence, capability, and correction tightly coupled.

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.

1,787 curated presets with measured parity, 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, enriched with 6 byte-reproducible semantic signals (magnitude, direction, certainty, significance) without calling an LLM. A hash-pinned ground-truth corpus of 16,146 events across ten benchmark pages ships as a release asset. Includes a CLI (@refract-org/cli), a Python SDK (refract-py), a web explorer, and an MCP server so an agent can query change streams directly.

refract-org/refract · Docs · AGPL-3.0

Models what you, your agents, and your machines can jointly do — across models, tools, machines, permissions, and human attention. Reads configurations across Claude Code, Cursor, OpenCode, Windsurf, Gemini CLI, Claude Desktop, and Codex CLI into one local graph, keeping what an assembled agent can do strictly separate from what it may do.

Click a node to inspect its dependencies, verified evidence, blast radius, and compound capabilities; simulate an outage to see what stops working. Distributable as a CLI via Homebrew (brew install zz-plant/tap/ambit), and 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 — generating meeting-ready operating postures in under 60 seconds with stop and reopen conditions written down before they're needed.

Deterministic: the same inputs produce the same call. Every posture carries explicit trip conditions, reopen rules, and boldness budgets. Ships with an MCP and Agent API so autonomous systems can query operating guardrails programmatically.

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.

265 essays so far, on a custom zero-dependency Cloudflare stack (Workers, D1, R2, newsletter delivery orchestration, and concept graph) 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. Ten proposed standards (including STD-08 Delegation and STD-09 Agent Chains), the Twelve Laws of Delegated Intelligence, operational mechanisms (kill switches, appeal paths, authority registers), and interactive diagnostics (Delegation Audit, Burden Modeler). 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