Data quality
Schema changes, missingness, out-of-range values, unseen categories, volume, and freshness. The quiet upstream faults that degrade a model long before anyone suspects the model itself.
A free, automated health scan of one production AI model. We compare a recent production window against a known-good reference and hand back a clear red, amber, or green across drift, data quality, output shift, and governance readiness – read and contextualised by a senior engineer, in days, no labels required.
A model is a snapshot of the world at training time, and the world keeps moving. Accuracy decays quietly – nothing errors, nothing alerts – while infrastructure monitoring watches the box the model runs in, which is precisely the layer that does not degrade. Most teams find out from a complaint, not a chart. The Health Scan makes the invisible visible on one model, in hard numbers, before you commit to anything.
Every scan reads the same four layers of model health, using standard, checkable statistics rather than proprietary scoring – the same toolkit we instrument properly inside paid engagements.
Schema changes, missingness, out-of-range values, unseen categories, volume, and freshness. The quiet upstream faults that degrade a model long before anyone suspects the model itself.
Per-feature distribution shift between your reference window and current production, measured with standard statistics (PSI, KS, Wasserstein, JS divergence) and reported feature by feature.
Shift in the model’s own prediction distribution. This is the early-warning layer that works without ground-truth labels – which is the normal production case – so degradation shows up before outcomes confirm it.
A scored checklist of the evidence a risk function would ask for: prediction logging, a retained baseline, a model card, evaluation history, and automated alerting. If accuracy happens to be measurable from your data, we score that too.
The engine is automated; the wrapper is light process. Your total time investment is one short call and two data extracts.
A senior engineer comes back within one working day. Together we pick the one model to scan and agree the reference window – the training set or a period you know was good. A short data-handling note goes in place before anything moves.
You provide two extracts: the reference window and a recent production window, plus – if you log them – a column of the model’s own scores. No labels needed, no access to the model, no customer identifiers required.
The engine runs the statistics in minutes. A senior engineer then reads the result in context, sanity-checks it against false alarms, and writes the plain-English recommendation. Automation does the watching; judgement does the reading.
You receive the scorecard – overall status, a headline in plain English, per-feature drift, data-quality findings, and the governance checklist – and, if you want it, a short call with the engineer who read your results.
One page that answers the question. An overall red, amber, or green; a headline that says what moved and by how much; a per-feature drift table; the data-quality findings; the governance checklist; and a recommendation you can act on. Written to be read by the person who owns the risk, not just the person who built the model.
The Health Scan is a first-look health indicator on one model against one reference window. It does not confirm root cause, cover multiple models, or stand up ongoing monitoring – that depth is the job of the fixed-fee Onboarding & Assurance Audit, and the ongoing accountability is Managed ModelOps.
If the scan comes back green, you have earned a good morning’s reassurance and we will tell you exactly that. If it comes back red or amber, you will know precisely where – and what we would do about it.
Tell us a little about the system and we’ll come back within one working day to agree scope. No sequence, no campaign – a senior engineer replies.
Common questions we get from senior technology leaders evaluating this work. Direct answers, no hedging. Open one to read in full.
Two extracts from the one model we agree to scan: a reference window – the training set or a period you know was good – and a recent production window. If you log the model’s own scores or predictions, include that column; it powers the output-drift check. No labels are required, and we never need access to the model itself.
A short data-handling agreement goes in place before anything moves, and by default we work in your environment or on an extract you control. The scan reads feature and prediction columns, not customer identifiers – strip anything you would rather not share and the results are unaffected.
A senior engineer replies within one working day to agree the model and the reference window. Once the two extracts land, the scan itself runs in minutes and the read-through takes little longer – you typically have the report within days, not weeks.
You get the report and, if you want it, a short walkthrough with the engineer who read your results. A red or amber result usually points to the fixed-fee Onboarding & Assurance Audit, which instruments monitoring properly and confirms whether the drift is affecting decisions. A green result is a good morning’s reassurance – and a conversation about keeping it that way. No obligation either way.
Because the scan is almost entirely automated – the statistics run in minutes and a senior engineer spends a short time reading them in context. It is deliberately a first look at one model against one reference window, not a full assurance review: root cause, multiple models, and standing monitoring are the audit’s job. The scan exists to show you, in hard numbers, whether that conversation is worth having.
The scan is free, the report is yours either way, and the walkthrough comes with no pitch attached. The worst outcome is a good night’s sleep.