Build systems

A knowledge assistant built on your own documents: proposals, reports, methodology, archives, earlier research. Ask something, and the answer comes from your own sources, with the location included. If the answer isn't there, the system says so, instead of making something up.

Searching, summarizing, and surfacing material the organization already has. The proposal that's three-quarters already been written before now comes out of memory with citations, ready to be adjusted and approved. Every action is logged: which source, which answer, who signed off.

An organisation's knowledge sits in hundreds of documents and in the heads of three people. Working out what already exists, every single time, costs days, and when a colleague leaves, memory disappears with them. This system makes the past searchable without losing control: the model invents nothing, because it may only answer from your own sources. For consultancies that's the pile of proposals, for publishers the archive, for research groups the earlier work.

A document processor that reads, organizes, tags, summarizes, or converts material into another format. With a human who signs off before anything counts.

Speeding up the repetitive work around material. For curriculum developers: relinking material to objectives and metadata at a revision, presented as a proposal, confirmed by the editor. For editorial teams and publishers: one piece converted into the formats of different channels, for approval. Every proposal shows what it's based on; nothing goes out automatically.

Making material is the craft. Retagging, rearranging, and converting it is busywork that recurs with every revision and every channel, and that's exactly the work that eats the weeks. The pipeline takes over the preparatory part and leaves the judgment where it belongs: with the editor or developer who signs off. Faster where it can be, checkable where it must be.

A system that handles recurring questions and requests on its own, and stops the moment a decision is needed. No commitment counts until a human has approved it.

Answering the questions that come back every week (opening hours, procedures, standard requests, scheduling) from a managed knowledge base, with citations. Anything that deviates, is sensitive, or touches money goes to a human, with the preparatory work already done. And if the answer isn't in the knowledge base, the system says so, honestly.

Recurring questions are small one by one and a full day's work together. Outsourcing them to a standard chatbot is risky: a bot that speaks on the organization's behalf and makes the wrong commitments gets held to those commitments. There's case law on that by now. This front office is therefore bounded: routine handled on its own, decisions to the human, and a log of every answer. The time savings without the reputational risk.

Custom-built for the busywork that doesn't fit a standard product, on the same foundation: Trail, Retrieval, Approval, Constraints, Expertise, Storage.

Sometimes the scan points to a piece that's too specific for a standard pipeline: a valuation engine, assessment preparation, a link between systems. Then we build that, but never without an analysis up front, because building a system and then looking for what it's for is exactly how failed AI projects begin. Hence the fixed rule: first the scan, then the custom build.