AI and data
Models and pipelines that earn their keep, with evaluation you can audit.
Start your buildWhat's included
The work, and what you receive
Each capability names an artefact you end up owning, not an activity we perform.
Data pipelines
Tested, scheduled and monitored, with data quality checks that fail loudly.
You receive
- Pipelines
- Quality checks
- Monitoring
Evaluation suites
A fixed evaluation that every model version runs against, so a change is measurable rather than asserted.
You receive
- Evaluation suite
- Version comparison reports
Auditable decisions
A stored decision record per prediction, retained for the statutory period and readable without an engineer.
You receive
- Decision records
- Retention policy
Model deployment
Versioned rollout with the ability to roll back to a known-good model.
You receive
- Model registry
- Rollback procedure
Bias and fairness review
Measured across the segments that matter to your regulator, and reported whether or not the result flatters us.
You receive
- Fairness report
- Segment breakdown
Our approach
How this runs, and roughly when
Durations are bands, not promises. We revise them in the open when the work argues otherwise.
- 1
Establish the baseline
1–2 weeksWhat accuracy is worth, and what the current process achieves. Without it, a model has nothing to beat.
- Baseline measurement
- Success criteria
- 2
Build the evaluation first
1–2 weeksThe evaluation suite comes before the model, so every subsequent change is measured against a fixed target.
- Evaluation suite
- Held-out datasets
- 3
Model and iterate
4–12 weeksVersioned experiments with results recorded, including the ones that did not work.
- Model versions
- Experiment log
- 4
Ship with an audit trail
2–3 weeksDeployment, decision logging and the documentation a regulator would ask for.
- Deployed model
- Audit trail
- Model card
Tools we use
Boring, well-supported, and replaceable
We tell you when something newer is worth its risk, and when it is not.
- Python
- PyTorch
- MLflow
- dbt
- Airflow
- Great Expectations
Related work
Where we have done this
In-house product
HR technology / careersTalentrah: an AI career platform built for the market it serves
An AI career platform for Nigerian job seekers, built around a constraint US-first competitors do not have: the people who need it most cannot afford $24 a month.
1,136 Automated tests
Questions
Asked often enough to answer here
Working against a constraint?
Tell us what it is. We will tell you what we would build, what we would not, and what it would take.