Eighteen dimensions per solution — the problem, the buyer, the architecture, the data flow, the standards, the evidence, how we will produce the first real number, the regulatory and liability posture, the economics and price, the competition, the limitations and the roadmap. Including what is not built.
The cross-cutting truth that no single solution dossier can see. Every claim below was verified by direct probe of the production API and the public Studio page source on 2026-09-01, or is explicitly labelled as company-reported. Where this page conflicts with a solution dossier or with the marketing site, this page is correct.
Ardia Precision Health is one product, four engine paths, and eleven names.
The product is a governed retrieval-and-drafting loop: text comes in, a regex de-identifier strips structured identifiers, a shared safety preamble plus a short persona prompt (60 to 250 words, depending on the persona) wraps the query, a retriever pulls real citations from PubMed and a hand-curated set of CMS Local Coverage Determinations, one Google Gemini call produces an answer, and six deterministic gates inspect that answer and withhold it entirely if any gate fails. That loop is real. It runs. Anyone can probe it from the outside without credentials.
Everything else in the portfolio is a label on that loop, with two exceptions.
The routing is one line in studio.html:
```
ENGINE_MODEL = { molec:'molec', toxiq:'molec', pulmo:'tara', meridian:'tara', aria:'aria', lumen:'lumen' }
```
Ten named models on the site collapse to four server-side prompts. ToxIQ is MolecuIQ — same system prompt, same LCD corpus, same call; the only difference between running them is which button the user pressed. PulmoIQ is TARA. Meridian, in the Studio, is TARA. The site's separately listed "MolecuIQ — genomics and pharmacogenomics" under Precision Medicine is the same molec revenue-cycle prompt, which contains zero instructions about variant interpretation, star-allele calling, NCCN or CPIC. Three advertised models are one artifact.
This is verifiable in about ten minutes by anyone who views source. Posting model:"pulmo" directly to /api/run returns {"error":"bad_model"}. The UI works only because the browser rewrites pulmo → tara before sending. A diligence reader who finds that rewrite next to models.html's "deterministic Python engine — no LLM" label will not stop at the routing; they will re-open every other claim on the site.
Two further corrections that must travel with any portfolio description:
The engine is Google Gemini, not Claude. GET /api/run returns {"ok":true,"provider":"gemini","gated":false}. Fast tier serves gemini-flash-lite-latest (~3s); Scholar tier serves gemini-flash-latest (~54s). A dormant call_anthropic() path exists in the code but Gemini is checked first and always wins. Both model IDs are unpinned -latest aliases.
Imaging does not work and is disabled twice. The API returns {"error":"uploads_disabled"} for any attachment, and studio.html independently hardcodes attachments:[], so a file the user selects is read to base64 and then discarded. Lumen explains imaging report text. It has never processed an image. Any interface element implying otherwise should be removed, not relabelled.
The two exceptions to "it's all one loop":
models/meridian/clfs.py contains genuinely deterministic CLFS/PAMA compounding arithmetic, unit-tested (34/34 passing, company-reported), mirrored by a client-side calculator on model-pama.html. It is not wired into the Studio answer path. A buyer meets two different Meridians depending on which page they land on — a real calculator on the marketing page, and Gemini freehanding PAMA answers as TARA in the product demo.cadence_model.joblib. The Cadence a customer can touch is not the Cadence that has a number.The honest one-sentence version: Ardia has one product, one deterministic module, and one trained model, instantiated as eleven go-to-market names. That is not a demotion. It is the only defensible strategy in the set — a solo founder who built one governed reasoning surface and packaged it eleven ways has done something sensible. The failure is not the architecture. The failure is describing four prompts as ten models to people who can read JavaScript.
The moat is not the model. The model is a third-party API any competitor can call with a credit card and forty lines of Python. The moat, to the extent one exists, is three things:
The corpus. A hand-curated set of CMS Local Coverage Determinations, returned with working cms.gov links (verified: L35025, L38045). This is real grounding, and it is the only asset in the company that a competitor cannot replicate by calling an API. Its honest size: eight LCDs, company-reported, curated by one engineer, covering molecular and urine-drug-testing coverage. That is a weekend of work for a domain expert and a genuine asset for a lab that does not have one. It is not eight years of proprietary claims data.
The gate layer. Six deterministic gates — non_diagnostic, safety_escalation, scope_of_practice, de_identification, honesty, human_in_the_loop — that return a reason on every call and, on failure, withhold the answer entirely. This is enforced in code and verified on production. Cite-or-abstain and policy-override are enforced separately in retrieval and answer-binding and are not counted among the six. A gate layer that actually refuses to emit is more than most healthcare AI demos have. It is also six regexes.
The audit trail. Every call carries gate results with reasons. That is the substrate of a compliance story, and it is the right thing to have built first.
Now the honest accounting of the moat's size. Most specialty corpora are unbuilt. There is no curated GOLD/GINA guideline corpus — pulmonary queries are grounded in retrieved PubMed literature, and that retrieval genuinely works (a GOLD/COPD probe returned PMID 40050074 and PMID 38032494, both real, both linking). The earlier internal claim that pulmonary returns zero sources is false and must never be repeated. But "grounded in retrieved PubMed literature" is a materially weaker claim than "grounded in the GOLD guidelines," and only the first one is true. There is no NCCI/MUE data, no CLFS rate table in the answer path, no CPIC content reaching the precision-medicine prompt. Crucible and Sentinel are modelled targets. The moat is one corpus, six gates, and one engineer's judgment — defensible against a generic chatbot, not yet defensible against a funded competitor who hires two coders and buys a policy feed.
And there is a structural crack in the gate layer that must be stated at portfolio level because it gates the entire company: the de_identification output gate calls the same deidentify() function used on the input. It is not independent verification. It is the same regex asked the same question twice. Anything the input pass cannot see, the output gate is mathematically guaranteed not to see. On probe, a prompt containing a plain personal name produced an answer containing that name, while the gate returned passed with the reason "no PHI identifiers present." Structured identifiers — dashed SSN, delimited phone, email, labelled MRN, member ID, US-format dates, context-gated ZIP — do redact. Plain personal names do not.
That single gap is the highest-leverage item in the company. Sentinel's output does not meet 45 CFR 164.514(b) Safe Harbor, so it remains PHI. Which means Ardia cannot honestly sign a BAA. Which means no pilot can touch real data. Which means no reference customer. Which means no revenue. Every zero in the company traces back to one regex that does not detect names.
Picture a Tuesday morning that has not happened yet.
Google rolls gemini-flash-latest forward. Ardia does not pin model IDs — it resolves the -latest alias at call time — so nobody at Ardia is notified, no deploy happens, no commit lands, and no log line changes. The new model is slightly more verbose and slightly more inclined to hedge. It now begins medical-necessity answers with a paragraph of context before the citation.
The honesty gate is a regex looking for cited claims in a particular position. It starts failing. Because a failed gate withholds the answer entirely, the product now returns nothing.
Not for one product. pulmo → tara and meridian → tara and tara itself all break at the same instant. On the other path, toxiq → molec and the "precision medicine" listing and MolecuIQ all break together. Six of the ten advertised models go dark from two prompt paths, in production, silently. There is no eval harness anywhere in the company that would detect it. There is no per-persona regression suite. There is no canary, no golden-answer set, no accuracy metric of any kind to degrade against — you cannot detect a regression from a baseline you never measured.
The same blast radius applies to a one-word edit to the shared GUARDRAIL preamble, and to a single character in any of the six gate regexes. Lumen and Aria hold their own engine keys, and that protects them from a TARA prompt edit — but not from a preamble edit, a gate edit, a Sentinel edit, or a model roll, because everything except their ~60-word and ~250-word persona strings is shared. Lumen's dedicated key is a real distinction and an oversellable one: a dedicated engine key is not a dedicated model.
Three concrete, cheap remediations, in order:
-latest aliases with explicit versioned IDs. This is a one-line change and it converts an unmanaged reproducibility risk into a managed upgrade decision. Until it ships, Ardia cannot honestly claim reproducibility to any regulated buyer, and cannot answer "what model produced this appeal letter" for any answer older than the last silent roll.Until item 1 ships, the honest disclosure to any investor or buyer is: the served model can change without our knowledge, and we would find out from a customer.
MolecuIQ, ToxIQ, and Meridian are not three products. They are one product sold to one buyer under one contract.
The buyer is a single person: the CFO or revenue-cycle director of an independent molecular and toxicology lab facing the 2027 PAMA rate cliff. That person owns the denial problem (MolecuIQ), the toxicology denial problem (ToxIQ — literally the same prompt), and the rate-cliff planning problem (Meridian). One budget. One security review. One BAA. One implementation. Two of the three are the identical system prompt.
Selling these as three products triples the sales motion, triples the security review, triples the objection surface, and produces exactly one contract. It also creates the diligence problem in Section 1 for free: the moment a prospect's technical person compares the three, they discover the routing map, and the conversation stops being about denials.
The land-and-expand sequence nobody has written:
Stage 0 — Prerequisite, not a stage. Sentinel detects personal names. Model IDs pinned. Golden set running. Until these three are done, no pilot conversation should start, because the honest answer to "can you sign a BAA" is no, and discovering that in week three of a sales cycle is worse than not starting.
Stage 1 — Land on the cliff, not on the appeal. Meridian's calculator is the wedge, and it should be the real calculator — clfs.py, or the client-side mirror, both of which genuinely work. A lab CFO will run a free rate-cliff model against their own test mix because the 2027 cliff is a board-level problem they already have. This is a deterministic arithmetic tool. It requires no BAA, no PHI, no accuracy claim, and no AI story. It is the only asset in the portfolio that can be given away at zero risk and still create urgency. Fix the two-Meridians problem by making the calculator the product and the Studio persona a supporting explainer, not the reverse.
Stage 2 — Convert the model into a diagnosis. The calculator output is a number: dollars at risk across 2027–2029. The follow-up is a free denial audit — the lab pastes 20 de-identified denials, MolecuIQ classifies them and drafts appeals against the LCD corpus. This is where the honest framing does the selling: we refuse to draft an appeal the record does not support, and we withhold any answer that fails a gate. Every competing tool over-promises here; the refusal is the differentiator. Twenty denials is small enough to do under a limited DUA, or on synthetic data if the BAA is not yet signed.
Stage 3 — One paid pilot, one BAA, one scope. Molecular denials only. Not toxicology, not pulmonary, not genomics. One code family, one payer, one workflow, ninety days, with a measured outcome defined before the pilot starts: appeal acceptance rate versus their historical baseline. Ardia currently has zero measured outcomes of any kind. The first pilot's only real product is the first number.
Stage 4 — Expand within the same contract, not into a new sale. Toxicology denials are a scope amendment, not a new deal — it is the same prompt, the same corpus family, the same buyer, the same BAA. Add the three urine-drug-testing LCDs to the curated set and turn it on. This is the cheapest revenue expansion in the portfolio and it costs one line in a statement of work.
Stage 5 — Only now, a second corpus. With one reference customer and one measured number, the LCD corpus can be expanded deliberately toward whichever code families that customer actually loses money on. Corpus depth is the moat; build it against real denials, not speculatively.
Note what is absent from this sequence: pulmonary, elder care, imaging, genomics, and the word "platform." Those are Stage 6 and beyond, and putting them in the deck at Stage 1 is what turns a credible RCM sale into an unfalsifiable AI pitch.
One immediate correction: as observed, the /solutions ToxIQ card described MolDX Z-code registration and NGS tiers — MolecuIQ's domain, not toxicology's. Verify against the live page, since copy is being corrected, but this is the visible symptom of the underlying fact that the two products are one prompt.
Aria, Ardia One, and Cadence do not belong in the same company as MolecuIQ, ToxIQ, and Meridian. Not because they are bad — because every dimension that determines how a company is built points the other way.
Different buyer. Lab RCM sells to a CFO with a P&L problem and a budget line for revenue recovery, with a sales cycle measured in weeks and a value proposition denominated in recovered dollars. Elder care sells to a family, a home-care agency, a health plan, or a health system — each with a different budget, a different procurement path, and none of them with a "recovered dollars" number to point at. There is no overlap in the call list.
Different regulation. Lab RCM is administrative work on claims. The non_diagnostic posture is easy to hold because nothing in a denial appeal is a clinical claim. Elder care puts a conversational agent in front of a vulnerable person who may describe chest pain. Aria escalates suspected emergencies to 911 — in three self-selected single-run probes it responded with a 911-first instruction — but escalation reliability is entirely unmeasured: no sensitivity, no specificity, no false-negative rate, no run-to-run consistency figure. A missed escalation in Aria is a fundamentally different category of harm than a bad appeal letter, and it lives one regex away from FDA SaMD territory that the rest of the portfolio comfortably avoids.
Different escalation path. The lab product's failure mode is "the answer was withheld" or "the appeal was weak" — recoverable, reviewable, human-in-the-loop by design. The elder-care product's failure mode is that nobody was called. Ardia has no notification infrastructure: Aria persists nothing, notifies nobody, dials nothing, and exchanges no data with Cadence or Ardia One. The triad is a modelled target with no interface, no shared state, and no code.
Different evidence bar. Lab RCM can be sold on a workflow demo and a pilot number. Elder care, sold to anyone institutional, needs safety evidence Ardia has not begun to generate — and Cadence's 95.45% is not it. Cadence is a phone-motion activity classifier on a public benchmark with zero older adults ever scored and no feature-extraction code connecting it to a real sensor. It must never be cited as evidence for anything Aria does; they are different models, and one of them has no number at all.
What this means for focus. Elder care is a second company that happens to share an engine. The right move is not to kill it — Cadence is one of only two measured things Ardia owns, and the elder-care thesis is genuinely large — but to stop selling it in the same breath. Concretely: keep elder care as a labelled research line with no revenue expectation, no pilot pursuit, and no space in the lab-RCM deck, until the lab SKU has one paying customer. A solo engineer running two go-to-market motions runs neither.
Every Ardia framing must acknowledge the five pillars. Here is what is genuinely built under each, versus labelled.
Pillar 1 — Precision medicine. Labelled. The site lists a precision-medicine model. It resolves to the molec revenue-cycle prompt, which opens "You are MolecuIQ, Ardia's molecular/toxicology denial-recovery assistant… This is ADMINISTRATIVE revenue-cycle work" and contains no instruction about NGS interpretation, star alleles, NCCN or CPIC. One probe returned an answer consistent with the CPIC clopidogrel guideline; that is n=1, self-graded, from a pipeline that supplies the model no CPIC content. It demonstrates Gemini's fluency, not Ardia's grounding.
Pillar 2 — Molecular and genomic diagnostics. Genuinely built, at the reimbursement layer only. Real CMS LCDs with working links, a real denial-classification and appeal-drafting loop, real refusal behavior. This is the strongest thing in the portfolio. It is reimbursement reasoning, not diagnostics. The kidney SKU sits here and is not built.
Pillar 3 — The 2027 PAMA rate cliff. Genuinely built as arithmetic, not wired into the product. clfs.py is real, deterministic, and unit-tested. It is not a CLFS rate engine — it applies a user-chosen annual cut to a user-supplied test mix — and it does not execute in the Studio. Two Meridians. Fix this before anyone diligences it.
Pillar 4 — Pulmonary and respiratory care. Labelled, with real but generic grounding. PulmoIQ is a browser-side label that routes to TARA. There is no pulmonary system prompt. Retrieval returns real PubMed literature — the citations work — but no curated GOLD/GINA corpus exists. The honest sentence is "grounded in retrieved PubMed literature; a curated GOLD/GINA corpus is not built." Note also that PFTs are paid under the Physician Fee Schedule and OPPS, not the CLFS, so pulmonary has no actual PAMA exposure despite sharing Meridian's engine path.
Pillar 5 — Elder care. One measured classifier that ships in nothing, one persona that ships with no measurement. Cadence's number is real and company-reported; the shipped Cadence surface never loads it. Aria is ~60 words of persona text at server.py:79-82. Ardia One is a modelled target. No integration exists between any two of the three.
Toxicology, on the roster alongside the five: it is MolecuIQ's prompt with a different button.
Portfolio-level status ledger. Two things are measured, both company-reported and neither independently reproduced: Cadence's 95.45% and Meridian's unit-tested arithmetic. Everything else is either live demo — it runs, with no published accuracy of any kind — or modelled target — not built. Crucible and Sentinel are modelled targets. There is no benchmark, no eval set, and no accuracy figure for any persona.
The honest zeros, which must never be softened. 0 customers. 0 pilots. 0 signed BAAs. 0 signed DUAs. $0 revenue. $0 raised. No real patient data has ever been processed anywhere in the platform. No clinical outcomes. Pre-revenue, Dallas-Fort Worth, founded December 2025.
Posture. Administrative and decision-support software, non-diagnostic always, not FDA SaMD. Aria escalates suspected emergencies to 911. Texas SB 1188 data residency and TRAIGA apply. The HIPAA matrix is self-graded 2/15 — which is the correct, honest score for a company whose de-identifier does not detect names.
Team. The founder (sole engineer — 15+ years enterprise, 9+ healthcare IT: Cigna, Teladoc, ECFMG, UnitedHealth/Optum). Manasa Jampani (Co-Founder, 10+ years payer operations: UnitedHealth, Interwell, ECFMG). Paramesh Kurapati (CEO, 2026). Sireesha Mamillapalli PhD, Geisinger Commonwealth — board, scientific oversight. One engineer wrote everything described on this page. That is the correct lens for reading both the achievement and the gaps.
The founder must decide whether Ardia is a lab revenue-defense company, and then delete or demote everything that is not that.
Not sequence it. Not deprioritize it. Decide it, and make the site say it.
The evidence points one direction. The only real corpus is the LCD set. The only deterministic engine addresses the PAMA cliff. The only coherent single buyer is the independent molecular and toxicology lab CFO staring at 2027. The only sales motion that closes without safety evidence Ardia cannot yet produce is administrative revenue recovery. The two founders' combined nine-plus and ten-plus years are in payer operations and healthcare IT — exactly this buyer. Everything in the portfolio that is genuinely built serves this thesis, and everything that is only labelled serves a different one.
The cost of not deciding is not that elder care progresses slowly. It is that the breadth itself is now the primary liability. Eleven names over four prompts is what forces the founder to defend an architecture instead of selling a workflow; it is what puts "deterministic Python engine — no LLM" on one page and meridian:'tara' on another; it is what makes a ten-minute source read recolour every honest claim on the site. The company is being harmed by its own surface area.
What deciding looks like, concretely:
clfs.py or stops answering rate questions.If the founder decides the other way — that elder care is the company — the same discipline applies in mirror image, and the lab products become the research line. What cannot work is the current state: two companies, eleven names, one engineer, four prompts, zero customers, and a routing map anyone can read.
The engineering is more honest than the marketing. Fixing that gap costs nothing and is the highest-return work available today.
TARA is Ardia's governed reasoning path — a shared safety preamble plus a roughly 90-word persona prompt wrapped around a single Google Gemini call, with regex de-identification in front of it (structured identifiers redact; plain personal names verifiably do not) and six deterministic gat
A deterministic, unit-tested compounding model that applies a user-chosen annual Clinical Lab Fee Schedule cut across 2027-2029 to a lab's own test mix to show cumulative revenue at risk — genuinely arithmetic, not yet a CLFS rate engine, and not wired into the product a buyer actually cli
MolecuIQ is a guardrailed assistant running on Google Gemini (gemini-flash-lite-latest, an unpinned "-latest" alias) — not Claude — that reads a pasted molecular-lab denial in free text, classifies it, is instructed by its system prompt to refuse appeals the record does not support, and dr
ToxIQ is a toxicology go-to-market label on MolecuIQ's engine path — the browser rewrites it to molec before the call — that reasons over drug-testing coverage, coding and medical necessity against a small hand-curated set of CMS urine-drug-testing LCDs and produces a cited draft appeal wh
PulmoIQ is a browser-side persona label for pulmonary and sleep reimbursement support that today routes to Ardia's shared TARA prompt on Google Gemini, grounded only in retrieved PubMed literature rather than any pulmonary coverage-policy or GOLD/GINA corpus; the framing is coherent and ho
Lumen is a roughly 250-word persona prompt over Google Gemini that turns pasted lab- and imaging-report text into a five-section plain-language explanation, wrapped in Sentinel — an in-development regex redactor that strips structured identifiers (SSN, phone, MRN, dates, ZIP) but does not
Aria is a warm, guardrailed conversational persona for older adults that today is a ~60-word system prompt over Google Gemini Flash Lite in a text box, with 0 customers, 0 pilots, 0 signed BAAs or DUAs, $0 revenue, $0 raised and no real patient data ever processed — it does not listen, spe
Cadence is two things sharing one name: an offline 561-feature scikit-learn logistic regression that labels 2.56-second windows of waist-worn phone motion into six activity classes at a company-reported, independently unreproduced 95.45% subject-independent accuracy on the public UCI HAR b
Sentinel is a dependency-free, pure-stdlib regex de-identifier that runs in front of every Ardia model and redacts a fixed list of delimited or label-anchored structured identifiers (dashed SSN, delimited phone, email, labelled MRN, member ID, US-format dates, context-gated ZIP); it has ze
Crucible is a small, deterministic, pure-regex guardrail layer that runs six named policy checks on every answer the Ardia Studio returns and withholds the answer entirely when one fails — enforcement verified in production, no measured sensitivity or specificity for any of the six gates,
Ardia One is the consumer/caregiver web application that puts Aria and Lumen prompt-framings, a Cadence-branded vitals mock-up (no Cadence model executes in this app), and an in-hub runner for the sibling models into a single elder-facing "care hub" — today it is one 113KB static HTML page