Medicoder Platform — AI Medical Coding Assistant

Status: under clinical-validation and compliance review. The system runs end-to-end against test notes; the no-PHI-persistence guarantee and the deterministic MDM scorer are now enforced by a hermetic CI test suite, but the substantive claims that follow — PHI handling end-to-end, MDM scoring fidelity, CPT code-selection accuracy — have not yet been independently audited, and nothing is deployed publicly. They are presented here as the project’s design intent and current implementation framing, not as audit-substantiated facts. The portfolio surface will tighten as each claim becomes independently verifiable.

A medical coding assistant that reads clinical encounter notes and recommends CPT E/M codes. Implements the 2025 Medical Decision Making complexity rules — analyzing problems addressed, data reviewed, and management risk using the CMS 2-of-3 scoring method.

What It Does

Paste a clinical note, select the encounter setting (office, inpatient, ED), and the system returns:

• A recommended CPT code, labelled a suggestion requiring coder review — no confidence score, because none is calibrated
• MDM complexity breakdown (problems, data, risk levels)
• Alternative codes with reasoning when documentation supports a different level
• Flags where the note is ambiguous about a specific MDM element and asks what was actually done — it never suggests adding content in order to reach a higher code.

Latency is not yet benchmarked; any speed figure is a target, not a measurement, and will be tied to a published benchmark before it becomes a load-bearing claim.

What’s Verified in Code Today

The scoring math is deterministic, not LLM-guessed: the model extracts evidence (problems addressed, data reviewed, risk), and the MDM complexity level + CPT code are derived by a deterministic scorer implementing the CMS 2-of-3 method. A committed benchmark (benchmarks/accuracy.json) runs that scorer over a 26-note synthetic golden set spanning five encounter settings — office new and established, ED, inpatient initial and subsequent — and complexity levels: 26/26 MDM-level match and 26/26 CPT match (100%). Honest scope — this validates the deterministic mapping given the evidence counts; end-to-end accuracy from raw note text (which adds the LLM extraction step) is the evaluation still pending below. The settings breakdown and the per-note results are readable in that committed file.

CI on main runs three gates: a committed-credential scan over every tracked file; a hermetic test suite covering the no-note-text-in-browser-storage gate, the deterministic MDM scorer, and the golden-note eval (a label-integrity self-check plus an accuracy floor); and a benchmark-freshness gate that fails the build if benchmarks/accuracy.json drifts from the engine.

One correction is worth stating plainly: the engine used to return a confidence score alongside each recommendation. It was a constant attached per code path — 0.96 on the Gemini path, 0.94 on Workers AI, 0.75 on the rule-based fallback — never derived from the note, the model, or any agreement measurement, so it was removed on 2026-08-03 and no score is published in its place. There is no calibrated number to publish; if one is ever added it will have to come from a measured agreement rate against the golden set, and be reported as such.

The repository is wider than the E/M coder described above: it also holds a claim scrubber, a denials-and-appeals module (CARC classifier and cited appeal letters), a patient cost estimator, per-route RBAC, a tamper-evident audit log, an OpenAPI spec with a drift gate, and clearinghouse and FHIR adapter interfaces with mocks. Those are committed code, not deployed or independently reviewed, and the same audit-pending framing below applies to all of them.

Compliance Posture (Design Intent — Audit Pending)

These framings describe the architecture as-implemented and (for note-text persistence) as CI-enforced; no independent third-party audit has yet verified them end-to-end against a deployed artifact. Treat the audit-level guarantee as pending until that audit exists.

• No note-text persistence (implemented and CI-enforced). Clinical notes are processed in-memory and not written to durable storage; a hermetic test drives the real frontend autosave classes with a marked synthetic note and fails the build if any note text reaches sessionStorage/localStorage. This is source/test verification — an independent code/infra audit against a deployed artifact is the next gating step.
• Session metadata only (CI-enforced). What persists is UI state (encounter setting, timestamp) in browser sessionStorage that clears on tab close — never note content; this is asserted by the PHI-persistence test in CI.
• Graceful degradation. When the LLM-assisted path is slow or unavailable, the system falls back to a rule-based path so users always get a response.
• Coding ethics: the engine codes from what is documented, and only what is documented. It does not invent diagnoses and does not suggest documentation changes whose purpose is a higher billed level.

What’s Needed For This Entry To Tighten

• An independent code/infrastructure audit confirming the no-PHI-persistence claim end-to-end against the deployed artifact (not just the source review), and/or
• A published evaluation of end-to-end CPT E/M coding accuracy from raw notes — i.e. including the LLM extraction step, going beyond the deterministic-scorer benchmark above — against a reference set, with the methodology and the dataset class disclosed (not patient-identifying data).

For inquiries (clinicians evaluating clinical workflow integration, healthcare ops teams, prospective auditors), contact us via the SUMequities contact page. Case study and technical deep-dive available on request for serious inquirers.