Data & AI
AIwhoseanswerscanbetraced,testedandgoverned.
We build data platforms, pipelines, analytical products, applied machine learning and agentic workflows with automatic lineage, evaluation floors, drift monitoring and recorded human control.
- Read/Summarise
- Draft/Propose
- Act with Authorisation
- Bounded Autonomy
- Read/Summarise
- Draft/Propose
- Act with Authorisation
- Bounded Autonomy
- Read/Summarise
- Draft/Propose
- Act with Authorisation
- Bounded Autonomy
Who this is for
The situation that makes this urgent
The mandate and boundaries below come directly from the approved service catalogue.
Executive metrics cannot be traced reliably back to source events.
Models are in production without a measured performance floor or drift response.
Agentic workflows are being considered for decisions with financial, legal or human consequence.
Three ways to engage
Assess, Transform, Operate
Each package has a defined duration, retained outputs, commercial model and objective exit criterion.
Assess
Data estate and AI readiness baseline.
Transform
Platform and applied AI build.
Operate
Managed data and model operations.
Capability detail
The technical substance, on the page and not behind a form
The operating capabilities and controls that define this practice.
Ingestion
Streaming, batch and schema contracts
Data platform
Storage, modelling and semantic layers
Applied AI
Evaluated machine-learning systems
Agents
Governed autonomous and semi-autonomous workflows
Lineage
Source event to published metric
Model operations
Drift, performance and cost monitoring
Technical standard
Requirements that apply to every engagement
Must and should requirements from the board-approved practice standard, made visible before a conversation.
Validate schemas at ingestion and quarantine invalid payloads with rejection reasons.
Generate lineage automatically from ingestion to every published metric.
Place every model behind an evaluation harness, performance floor and drift monitoring.
Enforce absolute client data isolation, including prompts, outputs and derivatives.
Record human review before consequential financial, legal, clinical, employment or regulatory actions.
Publish a model card for every production model.
Classify sensitivity and residency at ingestion and enforce placement technically.
Version data contracts with source owners and fail pipelines on breach.
Make published figures reproducible from versioned code, data and execution environment.
Report inference cost and cost per successful outcome quarterly.
How we work
Four phases. Objective exit criteria.
A gate passed on assertion has not been passed. One exit criterion per gate is shown here; the full set sits on each practice page.
Discover
Workload profile, dependency map and recovery objectives written down and signed off by the client owner.Design
Target architecture reviewed against the vendor-supported pattern for the engine in question; failure modes enumerated.Deliver
Recovery tested from backup within the stated RTO, evidenced by a timestamped test record.Operate
Runbooks in place, rota published, first quarterly service review held with measured figures.
The boundary
What we do not do
A capability list without explicit limits is only marketing.
Out of scope
Training foundation models from scratch or operating a research laboratory.
Never autonomous
Irreversible financial, legal, clinical or employment action without human authorisation.
Data boundary
No data brokerage or personal-data enrichment without evidenced lawful basis.
Evidence
The artefacts that make the claim testable
Evidence replaces unqualified promises and unattributed metrics.
Lineage graph
A queryable path from source event through transformation to metric or model output. Measurement method and review date accompany the delivered record.
Model and agent record
Evaluation results, known limits, drift state, permission rung and human approvals. Measurement method and review date accompany the delivered record.
Get started
A readiness baseline separates valuable, feasible AI work from demos that cannot survive production governance.
You keep the written baseline whether or not we continue together.