Generative AI, agents,and the data beneath.

Retrieval-grounded assistants and document work

Assistants that answer from your own corpus with citations, extraction pipelines that turn contracts, claims and invoices into structured records, and content systems that keep a house voice. We measure accuracy on a held-out set before anything reaches a user, and we design the refusal path as carefully as the answer path.

• Retrieval architecture, chunking and re-ranking
• Document understanding and structured extraction
• Evaluation harnesses, red-teaming and guardrails
• Prompt, model and inference cost optimisation

Agents scoped to one process, with a full audit trail

An agent is software with permissions, so we scope it like software with permissions. One process at a time, explicit tool boundaries, human approval where a mistake would be expensive, and a trace of every call it made. Most of our agent work replaces internal queues: reconciliation, triage, onboarding checks, back-office routing.

• Process mapping and automation candidacy review
• Tool and API integration with scoped credentials
• Human-in-the-loop checkpoints and escalation rules
• Observability, tracing and rollback

Governed data, tested pipelines, usable models

Lakehouse and warehouse builds, ingestion and transformation, a semantic layer the business agrees on, and reporting people actually open. Quality tests and lineage ship with the pipelines rather than arriving after the first incident, which is also what makes the AI work above defensible.

• Lakehouse and warehouse architecture
• Batch and streaming pipelines, dbt transformation
• Data quality, lineage, catalogue and access control
• Business intelligence, semantic models and forecasting