Large-scale occupational-health screening is one of those problems that looks simple from the outside and turns out to be anything but. Every contractor entering a high-risk worksite needs a medical-fitness evaluation, and at national scale that means tens of thousands of assessments a year — each one repetitive, time-pressured, and easy to apply slightly differently from the last. AMAN AI is a live platform, built by Crux Digits and partners, that speeds those evaluations up and makes them more consistent while keeping the clinician firmly in charge. It has been in production since February 2025 and has supported roughly 100,000 occupational-health assessment cases.
This is not a prototype or a benchmark. It is a real deployment running inside the existing clinical workflows of a national occupational-health provider, used for contractor recruitment and onboarding in the energy sector. Below is how the problem was framed, how the platform actually works, and how we would build something similar for you.
The problem: consistency at volume, without losing the clinician
National-scale fitness screening has two failure modes that pull against each other. Push for speed and throughput, and consistency suffers — different clinicians weigh the same borderline result differently, and subtle, latent risk gets overlooked in the rush. Slow down to be thorough, and the queue grows until onboarding becomes a bottleneck for the whole operation. The provider needed both: faster, more standardised evaluations and the assurance that latent risk would not slip through.
Two hard constraints shaped everything. First, the final decision had to stay with the clinician — this is occupational medicine, not an automated gate. Second, the solution could not require ripping out the workflows and systems the clinical teams already depended on. Any improvement had to slot into the way people already worked, not demand that they work a new way. That framing ruled out a black-box "AI verdict" from the start and pointed instead toward AI implementation designed around the people using it.
How the platform works
AMAN AI is a customised suite of AI models and agents, not a single model. It reads the same multi-source clinical inputs a clinician would, proposes structured, indicative outputs, and surfaces the reasoning behind them so the clinician can accept, adjust or override. Three decision-support models work together inside one platform:
- Classification AI — extracts and consolidates clinical parameters from unstructured, multi-source medical data, correlates anomalies against established guideline criteria, and proposes an indicative fitness classification the clinician can accept, modify or override.
- Cardiovascular-risk model — machine learning on vital signs and biomarkers surfaces latent cardiac risk as an interpretable, explainable risk score — including in people who look normal on standard indicators such as a resting ECG.
- Mental-fitness model (MHRQoL) — analyses physiological and biological markers to produce an indicative Mental Fitness Score that supplements, rather than replaces, established occupational-health guidelines.
The point of three separate models is that fitness is not one signal. Consolidating records, spotting cardiovascular risk, and assessing mental fitness are genuinely different problems, and treating them as such keeps each output focused and each piece of reasoning legible to the clinician reviewing it.
Why it can flag risk a standard test misses
The cardiovascular model is the clearest illustration of where machine learning adds something a checklist cannot. Rather than reading a single indicator in isolation, it learns correlations among indirect biomarkers and longitudinal physiological patterns. That lets it identify latent risk that may not be visible through standard indicators alone — and, critically, it reports that risk as an explainable score, not an opaque flag. The clinician sees why the model is concerned, which is what makes the signal usable in a real consultation. This kind of pattern-finding on messy, real-world clinical data is the heart of applied machine learning.
The approach: built into the workflow, not bolted onto it
The technical work was only ever half the job. The harder half was making three models behave as one dependable system inside a busy clinical environment. That meant ingesting messy, unstructured, multi-source data and turning it into consistent structured parameters; combining model predictions with guideline-based criteria so outputs stay aligned with established occupational-health standards; and presenting everything through the tools the clinical teams already use, so adoption did not depend on changing anyone's habits.
Delivering that reliably at national scale is squarely an application development problem as much as a data-science one: data pipelines, integration with existing systems, auditability, and an interface a clinician can trust under time pressure. The models are the clever part; the platform around them is what makes the cleverness safe to depend on every day.
Maker-checker: keeping the clinician in control
The governing principle of AMAN AI is a clinician-led maker-checker model, and it is the single most important design decision in the whole system. The AI is the "maker": it proposes an indicative classification, a cardiovascular-risk score and a mental-fitness score, each with supporting reasoning. The clinician is the "checker": they review, and they accept, modify or override every output. The system never makes the final fitness decision.
This matters for three reasons. It keeps accountability where it belongs — with a qualified clinician. It adds a layer of systematic, consistent AI validation on top of human judgement, so two safety nets catch what one might miss. And it makes the whole thing honest: every fitness, cardiovascular and mental-health output is explicitly indicative, supporting the clinician's decision rather than substituting for it. Clinical deployment of this kind also requires appropriate regulatory clearance and clinician oversight, and the platform is built to support that, not to bypass it. It is the same philosophy behind our other healthcare AI work — AI that earns a clinician's trust by being transparent and staying subordinate to their judgement.
Scale and impact
Since going live in February 2025, AMAN AI has supported roughly 100,000 occupational-health assessment cases in production. The figures below are delivered facts from a live deployment, not projections or industry benchmarks:
- ~100,000 occupational-health assessment cases processed.
- Live since February 2025, running inside the provider's existing clinical workflows.
- Three AI decision-support models — classification, cardiovascular risk and mental fitness — in a single platform.
At that volume, the value is compounding consistency. Every assessment gets the same systematic AI check before a clinician signs off, the same structured consolidation of records, and the same explainable risk scoring — which is exactly the kind of repeatable, traceable rigour that high-volume screening needs and manual review struggles to sustain across tens of thousands of cases.
Who this is for
AMAN AI was built for an occupational-health provider in the energy sector, but the pattern applies wherever high-volume, expert evaluation has to stay both fast and consistent without removing the human decision-maker. If you run large-scale medical-fitness screening, clinical-document triage, or any assessment workflow where a qualified professional must keep the final call, the maker-checker model translates directly. It is especially relevant to regulated, high-stakes settings — healthcare, occupational health, and safety-critical industries such as energy — where both throughput and defensibility matter, and where a black-box decision would simply not be acceptable.
How we would build something similar for you
We would not start by building three models. We would start with your workflow: where the volume is, where consistency slips, where latent risk hides, and where the clinician's judgement is non-negotiable. From there we would scope a focused pilot on your own data, prove the gain, and only then expand — the same honest, incremental approach you will find across our clinical NLP and ECG decision-support work. The maker-checker principle would carry over: AI as a fast, consistent, explainable second opinion, with your experts keeping control and accountability.
Transparent scoping is on our pricing page, and the fastest way to find out whether this approach fits your operation is to tell us what you are trying to improve. Get in touch and we will scope a focused pilot on your own data — and report real numbers, the same way AMAN AI does.
Live deployment: delivered by Crux Digits and partners at a national occupational-health provider, in production since February 2025, supporting large-scale medical-fitness evaluations for contractor recruitment and onboarding in the energy sector. Client name withheld pending permission. A clinician-led maker-checker model keeps every fitness, cardiovascular and mental-health output indicative — supporting, never replacing, the clinician's final decision; deployment requires appropriate regulatory clearance.
Frequently asked questions
Does AMAN AI make the final fitness decision?
No. It proposes an indicative classification, plus cardiovascular-risk and mental-fitness scores, that the clinician can accept, modify or override — a controlled maker-checker approach that keeps decisions clinician-led while adding systematic AI validation.
How can it flag cardiac risk when the ECG looks normal?
It learns correlations among indirect biomarkers and longitudinal physiological patterns, identifying latent risk that standard indicators may miss, and reports it as an explainable risk score the clinician can interpret rather than an opaque flag.
Does it fit existing clinical workflows?
Yes. It is a customised suite of models and agents designed to operate inside existing workflows and tools, aligned with established occupational-health guidelines, so clinical teams do not have to change how they already work.
Is AMAN AI a real, live deployment?
Yes. It has been in production since February 2025 at a national occupational-health provider in the energy sector and has supported roughly 100,000 assessment cases. The client name is withheld pending permission to publish it.