Data and AI platforms

Every stage 01 ends with a verdict on the data the use case depends on: it supports the case, it does not, or it will once a named problem is fixed. The verdict comes before anyone builds anything. Learning in week three that a key field is filled in a third of the time costs far less than learning it in month four of a build. Industry data shows the same risk: in February 2025 Gartner forecast that through 2026, organisations will abandon 60% of AI projects not supported by AI-ready data, from a global survey of 248 data management leaders.

When the data is the main risk, stage 01 includes the Data Readiness Review, a fixed-scope diagnostic of 1-3 weeks inside stage 01, not a stage of its own. It scores every source the use case depends on:

You receive a scorecard per source and a remediation plan for anything below the bar, with the fix and a rough effort. We work with the owner of each system of record, because reporting layers built on top often hide the gaps.

An answer drawn from your documents must show its source.

• Retrieve when the step needs it, and keep the rest of the context small.
• Cite or abstain: every answer points to the passage it rests on, and when nothing supports an answer the system says so and hands the question to a person.
• Quote or discard: in extraction, a claim needs a verbatim quote from the source, or it is thrown away.

Tenderix, our procurement tool, applies quote-or-discard to tender documents in German, French and Italian, and reports a missing score as missing. FoodHalo’s coach looks up nutrition figures through tools against the product database and may not state a figure it did not retrieve.

Many processes contain personal data that the model does not need to see. Our internal build replaces every name, address and identifier with a stable alias before content reaches a model; the same person keeps the same alias within a run, so the model can still reason about who did what. A separate check inspects everything about to leave and fails the run if it finds a value that should have been aliased. The aliases are reversed only at the end, locally, on your side.

Each task goes to the cheapest model that passes its evaluation, and budgets are set before the first request.

We report cost per completed task: the number that tells you whether the system still pays for itself.

• Start read-only. Write access comes one action at a time, each with its own approval path.
• Give every connector the smallest scope that does the job.
• Keep the decision path deterministic where it matters: in kredis, the decision endpoint other systems call reads precomputed state, with no language model in the path.
• Keep case history append-only, so any past state can be reconstructed.
• Isolate tenants, with row-level security enforced in the database.

kredis assembles counterparty files from GLEIF, Companies House and embedded OFAC, EU, UK and UN sanctions lists.

• WorkAnonymise before inference
• WorkA tiered document intelligence pipeline
• WorkA grounded knowledge copilot for field service
• WorkA multi-channel AI concierge, in production
• ProductsAtrium
• ProductsFoodHalo
• Productskredis
• ProductsTenderix