Industry
Build clinical AI a clinician can check in ten seconds
Healthcare systems fail on the boring parts — identity, terminology, consent and the audit trail — long before the model is the problem. Digiaeon builds clinical and operational AI that cites its source, abstains when unsure, and stays on the right side of the device boundary.
Healthcare & Life Sciences
Clinical and operational systems where being wrong is expensive — built with the citation attached
- Now
- 3 applications
- Next
- 2 applications
- Frontier
- 2 applications
- Constraints we design around
- 5
Context
Three forces reshaping this sector.
What changed recently enough that the old architecture is now the wrong one.
Ambient documentation became table stakes, so the value moved downstream
Every major EHR now ships or resells an ambient scribe, and the note itself is no longer a differentiator. The unsolved work sits either side of it: getting structured, codable facts out of the encounter, and driving the orders, referrals, authorisations and follow-ups that the note implies. That is integration and workflow engineering, not transcription.
Prior authorisation is being forced onto APIs on both sides of the wire
The CMS interoperability and prior authorisation rule pushes impacted payers onto FHIR-based CRD, DTR and PAS APIs with decision timelines attached, and India’s NHCX is pulling claims exchange in the same direction. Providers that still assemble evidence by fax and portal will feel the gap immediately — and payers that automate denial without automating the reasoning will feel the appeal volume.
The regulatory line around clinical decision support hardened
Between the FDA’s clinical decision support guidance and predetermined change control plans, the EU AI Act’s high-risk obligations, and India’s DPDP framework, the question is no longer whether AI is allowed in the clinic. It is whether a clinician can independently review the basis of the output — which is an architecture requirement, not a legal footnote.
Applications
What we would build here, in order.
Grouped by how close each one is to production today — not by how impressive it sounds in a board pack.
Build it this quarter
Current models, current data. The engineering is understood and the risk is schedule risk, not research risk.
Ambient documentation that produces structured facts, not just prose
Diarised ASR, speaker attribution, and a note generator constrained to what was actually said — then a second pass that emits FHIR Condition, Observation and MedicationStatement resources for the parts the chart can consume. The failure mode to engineer against is the hallucinated negative: a model writing “denies chest pain” when the topic never came up. Every assertion carries a transcript offset, and the clinician attests before anything is filed.
Prior authorisation evidence assembly
Parse the payer’s medical policy into machine-checkable criteria, retrieve the supporting evidence from the chart, and assemble a packet where every criterion points at a dated document. Submission rides X12 278 or FHIR PAS where the payer supports it, portal automation where it does not. The system never asserts a criterion is met without a citation — an unmet criterion is surfaced as a gap for staff to close, not quietly filled in.
Revenue cycle: denial prevention and coding assistance
A classifier over historical remittance advice learns which claims your payer mix denies and why, and flags them before submission rather than after. Alongside it, ICD-10-CM and CPT suggestion with NCCI edit checks and CDI query drafting — always as a suggestion a coder accepts or rejects. Autonomous code submission is where compliance exposure lives, so the human stays in the loop by design, not as a phase-two promise.
One thing has to change first
Buildable, but gated on something outside the model — labels, a consented holdout, an integration, or a process the business has to agree to change.
Patient navigation and triage with hard escalation rails
A front door grounded strictly on your own clinical content and coverage rules, answering in the languages your catchment actually speaks. Triage runs against a deterministic red-flag rule set that sits in front of the model, not behind it: chest pain, stroke symptoms, suicidal ideation and paediatric fever route straight to a human or an emergency number regardless of what the model would have said. Retrieval is scoped by the caller’s identity so the system cannot leak another patient’s record.
Imaging workflow orchestration — the worklist, not the read
Prioritising studies, reconciling DICOM metadata against HL7 ORU results, detecting unreported studies, routing priors, and closing the loop on incidental findings that fall through follow-up. This is high-value operational work that sits deliberately outside the diagnostic claim, which keeps the regulatory posture simple. If a project genuinely needs detection on the pixels, that is a device programme with a different budget and timeline, and we will say so.
Scope it as research, with a gate
Credible and worth a prototype. We would run it behind an evaluation gate and say plainly if it does not clear the bar.
Real-world safety signal detection
Individual case safety report intake from call centre transcripts, literature and complaint channels, with MedDRA coding assistance, duplicate detection against existing cases, and E2B(R3) generation for submission. The same extraction spine serves clinical trial document work — protocol digest, eligibility criteria structuring, source document abstraction. Precision matters more than recall for coding, recall matters more than precision for intake, and the two are tuned separately.
Capacity, flow and theatre scheduling under uncertainty
A learned arrival and length-of-stay model feeding a discrete-event simulation of beds, theatres and staffing, so a scheduling change is tested against a distribution of days rather than an average one. The honest constraint is that the optimiser’s answer is only as good as the operational data behind it — if theatre start times are recorded to the nearest half hour, the model cannot resolve fifteen-minute decisions, and we instrument before we optimise.
Constraints
What makes this sector different.
The part of the build that does not transfer from anywhere else. Every one of these has to be answered in the architecture, not in the project plan.
PHI minimisation has to survive free text
Structured de-identification is solved; narrative is not. A discharge summary that mentions the patient’s employer, their village, and the date of a distinctive procedure is re-identifiable even after the eighteen HIPAA Safe Harbor identifiers are stripped. We minimise at the boundary — send the span a task needs, not the record — pin inference to zero-retention, in-region endpoints under a signed agreement, and keep the de-identification decision documented because DPDP consent and HIPAA Safe Harbor are not the same test.
A wrong answer harms a person, so abstention is a feature
Most AI systems degrade gracefully. Clinical ones do not. Every deployment gets a hazard log in the ISO 14971 shape — what can go wrong, how likely, what detects it, what mitigates it — and an explicit “I do not know” path that routes to a human. The anti-pattern we refuse is the silent fallback, where retrieval returns nothing and the model answers from parametric memory anyway. No evidence means no answer.
Interoperability is where the schedule actually dies
HL7 v2 feeds arrive with site-specific Z-segments and a local dictionary nobody documented. FHIR R4 profiles diverge between US Core, IPS and ABDM. Epic and Oracle Health app review, scope approval and API throttling are measured in weeks. Terminology mapping across SNOMED CT, LOINC, ICD-10 and RxNorm is a standing programme, not a migration task. We ask for an integration sandbox and a real message sample in week one, because a project that gets them in month three has already slipped.
The software-as-a-medical-device boundary is an architecture decision
Non-device clinical decision support depends on the clinician being able to independently review the basis of a recommendation. That single clause decides whether you ship a product or run a device programme with a quality system, clinical evaluation and change control behind it. It is why our outputs carry citations rather than opaque scores, why we avoid time-critical alerting claims, and why we state the classification question at scoping instead of discovering it at procurement.
Population shift breaks models that still look accurate
A risk model trained on one payer mix, one season and one device fleet degrades quietly when any of those move. Discrimination can hold while calibration drifts, and falling prevalence collapses positive predictive value even with AUROC unchanged — the clinicians notice as a rise in false alarms long before a dashboard does. We ship subgroup-sliced calibration monitoring, a shadow-mode period before any model influences a decision, and a documented retraining trigger, accounting for the fact that outcome labels arrive weeks after the prediction.
Every one of these is discoverable in week one. Most programmes discover them in month three, after the architecture is already load-bearing.
Outcomes
What we design toward.
Targets and typical ranges for this class of system. Each one carries its basis, because a number without one is decoration.
- Documentation time returned per encounter
- 2–5 minDocumentation time returned per encounterPeer-reviewed ambient documentation evaluations cluster near the low end of this range; vendor-reported figures run higher. Treated as a hypothesis to baseline against your own note times per specialty, not a promise
- Unsupported statements in generated clinical text
- < 2%Unsupported statements in generated clinical textDesign target, measured against a clinician-adjudicated evaluation set before go-live and tracked on every model change
- PHI access attributable to a user, purpose and record
- Every callPHI access attributable to a user, purpose and recordDesign target enforced architecturally — the audit write sits in the request path, not beside it
- To a first production-safe release on a single workflow
- 10–14 weeksTo a first production-safe release on a single workflowTypical span for a scoped first cut on this stack, assuming an integration sandbox and named clinical owner from week one
These are design targets and published benchmark ranges, not delivered client results. We do not publish results attributed to clients.
Solutions
What we bring to this sector.
The practices that do most of the work here. Each one is the same discipline applied against the constraints above.
Blueprints
Reference architectures behind this work.
How the systems are actually assembled, with the engineering decisions and their costs stated. Reference designs, not case studies.
Grounded Answer Engine
A question-answering system over a company’s own documents that cites every claim to a source span, respects the permissions of the person asking, and refuses when the evidence is not there.
Read the blueprintDocument Intelligence Pipeline
A high-volume extraction pipeline that turns invoices, claims, KYC packs and trade documents into validated structured records, attaches a calibrated confidence to every field, and sends only the fields that genuinely need a person into an exception queue sized to the real error rate.
Read the blueprintQuestions
The awkward questions.
The ones a technology committee actually asks, answered the way we would answer them in the room.
Next step
Bring us the constraint you keep working around.
A 45-minute working session. We'll tell you what we'd build, what we'd not build, and roughly what it costs. No deck.