Ardia is an AI healthcare-infrastructure company building governed models for lab reimbursement, precision medicine and elder care. This is the single, honest answer to "what has Ardia actually built?" — each model, its AI engine, the standards it works to, and exactly how far along it is.
Eighteen dimensions per model — problem, buyer, architecture, data flow, standards, evidence, evaluation design, regulatory and liability posture, price and margin, competition, limitations and roadmap. Including what is not built.
Ten models across four focus areas plus a governed core. Each carries an honest status badge: ● measured = real numbers you can reproduce, ▶ live demo = runs today on a frontier model under guardrails (watchable, no published accuracy), ◐ modelled target = specified and being built, not yet measured.
Reads a toxicology lab's orders and claims and reasons about coverage, coding and medical necessity before a claim goes out — so definitive and presumptive drug-testing claims are clean the first time instead of denied.
The molecular-diagnostics counterpart to ToxIQ: maps molecular and NGS orders to the right CPT and MolDX Z-code, checks the governing LCD/NCD, and flags necessity gaps before submission.
Interprets next-generation-sequencing and pharmacogenomic results against published clinical guidance — matching NGS panels and PGx variants to the relevant NCCN and CPIC recommendations so a clinician sees the guideline-backed context. It is decision support, never a diagnosis.
Not an LLM — a deterministic engine that models the PAMA Clinical Lab Fee Schedule rate cuts and shows a lab exactly how much revenue is at risk as rates step down. The math is unit-tested so the numbers can never drift from the code, and it runs entirely in the browser with no PHI.
Focuses the reimbursement and documentation reasoning of the IQ family on pulmonary and respiratory care — pulmonary-function testing, spirometry and related services — checking coverage and coding before a claim is filed. Non-diagnostic.
Deep dive: the full PulmoIQ page →
A warm, voice-first companion for older adults — conversation, reminders and a gentle daily check-in. It listens for concerns, keeps a caregiver in the loop, and hands off to a human. It never diagnoses, and a suspected emergency is escalated to 911.
The one genuinely trained-and-measured model in the stack: a movement classifier that recognises daily activity from motion signals to give a picture of an older adult's routine. Trained from scratch and scored subject-independently on a public benchmark — the test people never appear in training.
Full numbers, confusion matrix and reproduction steps on Test Results → · the Cadence model page →
Turns a dense lab report or imaging summary into plain, calm language a patient can actually read — explaining what a result means and what to ask their clinician, always pointing back to the source report. It clarifies, it does not diagnose.
The reasoning core the persona models run on — a neuro-symbolic design pairing a frontier model's language reasoning with a symbolic policy engine that applies coverage and coding rules deterministically. It is administrative software, not an FDA medical device.
Deep dive: the full TARA page →
The privacy gate every other model sits behind — a deterministic rule engine that strips identifiers so no PHI reaches a prompt, enabling work on a Limited Data Set under a Data Use Agreement. Being built to the HIPAA Safe Harbor standard.
The evaluation harness that keeps everyone honest — the same subject-independent, reproducible-from-public-data protocol that produced Cadence's numbers is being generalised so every persona model gets benchmarked the same way, once real data is available under a BAA.
The persona models are not a single home-grown neural network — they run on frontier models under guardrails, with one classifier we trained ourselves. Here is exactly which tool does what, and how far each is proven.
| Tool | Vendor | Role in the stack | Status |
|---|---|---|---|
| Claude (Opus / Sonnet) | Anthropic | Under evaluation for the deliberative reasoning tier — clinical language and coverage reasoning. Not in the live path today. | Evaluated · not live |
| Gemini (Flash / Pro) | The reasoning engine actually running the persona models today (gemini-flash-lite) | In use · demo | |
| Cadence classifier | In-house (scikit-learn) | Trained movement / activity-recognition model — the one component with measured accuracy | Trained & measured · 95.45% |
| Open models | Open-source (Llama / Mistral class) | On-prem and cost-optimised inference for de-identified workloads | Roadmap |
Every persona listed on this page runs on Google Gemini under Sentinel's de-identification and TARA's policy guardrails. None has a published accuracy figure yet — that arrives model-by-model through Crucible, once measured on real data under a BAA.
What we're building (Phase 1): the ingestion-and-adjudication spine that turns a lab's raw claims traffic into governed, auditable decisions. Four pieces, built in order.
Ingests the standard claim (837) and remittance (835) EDI feeds so the system reads a lab's real reimbursement traffic — the raw material every downstream decision is built on.
Deterministic coverage logic: MolDX / DEX Z-codes, the governing LCD / NCD, and NCCI / UDT frequency rules applied as explicit rules — so a decision is explainable, not a black box.
Drafts a sourced appeal letter when a claim is wrongly denied, backed by a tamper-evident audit trail so every decision and its rationale can be reconstructed after the fact.
Will run on a HIPAA-eligible Google Cloud footprint under BAA — Vertex AI, Cloud Healthcare API, BigQuery, FHIR R4, pinned to us-central1 — on the path to SOC 2. Today's stack is Vercel + Gemini/Claude; the GCP footprint under BAA is on the roadmap.
Don't take the roster on faith — watch the models run, read the measured numbers, and try the deterministic math live.
