Pick the procedure. Dictate or paste the consult.
Dictum extracts the structured note in your template, with your vocabulary,
flags what payers will ask for, and writes the prior-auth packet.
Everything below the fold of this page used to be roadmap. Most of it shipped. Here's the short version of what's wired up at /ir right now.
The ask is simple: pick the right template before walking in — GAE, back pain, kypho, ESI, ablation, UFE, varicose veins — dictate your findings, and get a structured note out without fighting your EHR. Then have the prior auth letter waiting before the patient leaves the room.
Voice-to-structure is built and tested. 101 tests passing. Medical-grade speech recognition with your IR vocabulary loaded as keyterms — it knows kyphoplasty, basivertebral nerve, genicular artery, CEAP, Kellgren-Lawrence. Works on Android, Windows, and iPhone — no app install, runs in your browser.
Template selection — live. Pick from 9 procedure templates at /ir and the workspace pre-loads the schema, keyterms, and the gap rules that template needs. Sonnet 4.6 extracts the structured fields and cites the transcript span behind every value. The insurance defense layer — PA packet, necessity letter, P2P prep, denial appeal — is wired to the same extraction. One click per document.
Still tightening: per-procedure prompt tuning so the smaller templates don't miss obvious fields, and the dictation last-mile so the mic on /ir lands every transcript in the textarea without the user having to refocus.
Four steps between selecting the procedure and having a structured note.
Before you walk in, tap the procedure. Templates pre-load your IR vocabulary as Deepgram keyterms and set the output structure. One tap — SpineJack consult, GAE evaluation, UFE follow-up, whatever you're doing next.
After the visit, dictate what matters — on your phone, desktop, or wherever you document. Speak naturally using your own terminology. Dictum understands IR vocabulary out of the box and structures your dictation against the template fields. Type anything you prefer to type.
When the visit ends, the intelligence layer processes the raw transcript against your template. Free-form conversation becomes structured fields — HPI, relevant imaging, assessment, plan, procedure details. You didn't fill out a form. You had a conversation. This three-tier extraction pipeline is built and tested — real-time voice to structured clinical data.
Review the structured note. Tap to correct any misrecognitions — Dictum learns your corrections and adds them to future keyterm sets. Before you export, the insurance gap analysis checks your note against payer-specific criteria: is the KL grade documented? KOOS score? Compression therapy duration? Missing fields get flagged. One click generates a prior auth letter citing your documented criteria + guideline language. Export to clipboard, PDF, or FHIR.
The biggest risk with AI in clinical notes is invention. Dictum's extraction has a hard rule: if the model can't quote a transcript span that supports the value, it leaves the field empty. No clever guesses. No filler sentences a payer can pick apart.
Patient is a 68-year-old woman with right knee pain for three years. Failed PT, two HA injection series, and three steroid injections.
Weight-bearing X-rays show Kellgren-Lawrence grade 3 changes, medial compartment predominant. WOMAC pain score 14 of 20 at consult.
Plan is right GAE via left radial access, target descending genicular and superior medial/lateral genicular arteries, IPM microspheres 75–150 micron.
Each template ships a list of payer-specific checks. They run after extraction and tell you, in plain language, what a real denial reviewer would catch. Three tiers so you can ignore the noise and act on the real risks.
Eleven rules per template, nine templates. The dispatcher is generic — new templates plug in, no new code paths.
The extraction is the source. Every payer document below is generated from the same structured fields and the same citations. Edit once, regenerate any of the four.
The full bundle. Patient summary, conservative therapy timeline, imaging table, ICD/CPT codes, and payer-specific medical-necessity criteria with the documented response next to each one.
Single-page letter, doctor voice. The clinical rationale, the failed conservative therapies, and why this procedure for this patient. Drops into the EHR as a signed attachment.
The cheatsheet for the call. Most-likely reviewer objections with the documented response next to each one. Evidence citations ready to read out loud. Talking points by payer.
For when it gets denied anyway. Rebuts the payer's stated reason point-by-point with the same documented criteria, plus the published evidence relevant to that denial.
Each template packs the right vocabulary, output structure, and procedure-specific fields. Select before you walk in. The nine in the sidebar with a · dot are wired to the live extractor at /ir; SIJ and Facet/MBB/RFA are documented but not in the registry yet.
What it looks like when you dictate a GAE consult. Your words, your vocabulary, automatically mapped to template fields with structured extraction.
No vapor. Here's the stack, what's running, and where templates slot in.
Each card is a template with its vocabulary, output fields, and the terminology Dictum needs to nail. These are the exact keyterm sets that get injected into Deepgram.
Green is built and tested. Yellow is the active build. Gray is planned.
FastAPI proxy, WebSocket relay with asyncio.TaskGroup, Deepgram Nova-3 + Flux adapters, 8-type event schema, JWT auth, session persistence, rate limiting, parameter validation.
Web-based companion works on Android Chrome, Windows Chrome/Edge, and iPhone Safari. getUserMedia + AudioContext, QR pairing for walk-around dictation. Zero app install required. Audited: zero iOS-specific code branches.
9 IR templates wired in the registry (GAE, BVN, MILD, kypho, ESI/SNRB, UFE, renal ablation, cryo, varicose veins). Each has a Pydantic schema, ~11 gap rules, keyterms, and a JSON-schema tool spec for the extractor. Two more templates documented but not yet registered: Facet/MBB/RFA and SIJ.
Every structured field cites a transcript span (a prov_id). The extractor builds a paragraph-level catalog of the consult, then asks Sonnet to populate the schema with citations. Cited-but-unknown ids raise an error; uncited fields get nulled out instead of invented. This is the core hallucination guard.
Four documents generated from one extraction (PA packet, letter of medical necessity, peer-to-peer prep sheet, denial appeal letter). All four confirmed live against a real GAE extraction.
Eight of nine templates use a generic extraction prompt. Real-world test on kyphoplasty: the prompt missed target_levels=T8, acuity=acute, etiology=osteoporotic despite all three being verbatim in the dictation. Hand-tuned per-procedure prompts (already drafted in the spec markdowns) close the gap. Same priority: the mic on /ir reaches Sonnet and gets transcripts back, but the last hop into the textarea is timing-sensitive. The fix is in main.js; we're verifying behavior on real dictations.
PII/PHI redaction at the Deepgram level, audit logging, data retention controls, HIPAA documentation. Required before clinical deployment.
Structured notes in FHIR format. Integration patterns for Epic, Cerner. Copy-to-clipboard as the day-one bridge.
Not a generic transcription app hoping doctors will use it. A voice system designed around procedure-specific documentation — your procedures, your vocabulary, your templates. With an insurance defense layer that fights the denial before it happens. Purpose-built integration, not a feature checklist.
dictum — your voice, structured. your denial, defended.