Structured protocol
Every change trajectory is the same:
roadmap → workplan → session →
commit → memory note. The protocol is
codified in the repository's .claude/CLAUDE.md.
Capability · Operating model
AI-built today. AI-embedded next.
"AI-First" on this platform is not a marketing claim — it is the operational model. There is no team of human engineers writing code with AI as a sidekick. Humans write the roadmap and review the pull requests; the code itself is written by AI agents following a structured protocol.
234
feat + fix commits
2026-04-01 → 2026-05-20 · 50 business days · ~4.7/day
42
Daily prompt logs
2026-01-21 → 2026-05-20 · every AI session captured
5 in 8h
MFA platform shipped
Five ADM_02 sessions on 2026-05-14 — full MFA platform
6 in 21m
White-label brands
Four RCM_02 sessions on the same day, 23:25 → 23:46
Phase one
The platform is AI-built today.
A human owns the strategic roadmap. AI agents work each item to a Definition of Done — implementation, tests, security review, accessibility check, and commit. The cadence is not a sprint. It is the steady-state.
Per-session commits
Each session ends with a single commit — scoped, descriptive,
with a Co-Authored-By tag on the AI agent that
shipped the work. The commit message is the change log.
Mandatory review gates
Security checklist + accessibility checklist before "feature complete". The checklists are baked into the repo conventions — not policy, code.
Memory across sessions
Each shipped item generates a structured memory note that future sessions reference; the "context window" is the repository plus a curated index of prior decisions.
Verified cadence · 2026-04-01 → 2026-05-20
234 commits in 50 business days.
Hover or focus any day for the breakdown. Highlighted days sit above the noise; everything between is the steady-state cadence — ~4.7 commits per business day.
Total380
Days37
Avg / day10.3
Peak40
Apr 1Apr 22May 4May 14May 20
Hover a bar to see the date + commit count. Peaks carry a short note explaining what landed.
Phase two
The platform is AI-embedded next.
The same agentic primitives that build the platform will run inside it. A natural-language assistant for analysts; an AI Healthcare Analyst that drafts payer-policy rule changes from a PDF; an autonomous denial-resolution agent that proposes corrections, runs them dry, and submits for human approval.
Analyst copilot
Natural-language queries against tenant data, scoped to the user's RBAC. The output is structured rows + a re-runnable query — not a one-off result.
Healthcare Analyst Agent
PDF of a payer's revised policy in; YAML draft rule + dry-run feedback loop + human approval out. The agent never publishes; the human ships.
Denial-resolution agent
Per-CARC success rates drive a closed-loop agent that proposes corrections, runs them against a sample, and queues the approved ones for re-submission.
Substrate, not bolt-on
These features sit on the same canonical model and rules engine as the rest of the platform. The agents read and write through the same surfaces a human operator uses.
The competitive moat is the substrate, not any single AI feature.
Keep going