Services

AI Agents and RAG Systems That Cite Their Sources

Retrieval that cites. Agents that ask first. We build the full path: your data in, grounded answers out, and a person in the loop wherever the stakes are high.

How the pipeline works

  • Ingest: your PDFs, wikis and tickets, with access rules kept intact
  • Embed and index: chunked, embedded and stored in PostgreSQL with pgvector
  • Retrieve: only passages the user may see, ranked and cited
  • Reason with tools: an agent plans, calls your APIs and shows its steps
  • Human approval: anything irreversible waits for a person to say yes

What we build

Internal knowledge assistants, support and operations agents, document analysis and AI features inside existing products. Most AI demos stop at a chat box; we build the surrounding system: permissions, evaluation, logging and fallbacks.

Engineering discipline

Model calls run behind typed back-end services, outputs are validated, and failure modes such as timeouts and rate limits are handled deliberately.

Start small

A scoping sprint defines the use case, data sources and success measures before any build. Our AI Idea Validation Assistant can also pressure-test a new product idea first.

RAG and agent pipelineDocuments are ingested, embedded and indexed in Postgres with pgvector, relevant passages are retrieved with access rules applied, an agent reasons with tools, and a person approves irreversible actions.Ingestaccess kept01Embed, indexpgvector02Retrievecited03Reasontools, steps04Approvehuman gate05
Your data goes in, grounded and cited answers come out, and a person approves anything irreversible.

Who this is for

Organisations with a lot of documents and a lot of questions, support and operations teams drowning in repetitive work, and product teams that want AI features inside an existing application without losing control of correctness and cost.

RAG, agents and where each fits

PatternWhat it doesUse it when
RAG assistantAnswers questions from your documents, with sourcesPeople waste time searching knowledge scattered across files
Agent with toolsPlans steps and calls your APIs to get work doneA process spans systems and follows rules a person can check
Document analysisExtracts and classifies information from filesVolume makes manual reading slow or inconsistent

Guardrails we build in

  • Answers cite the passages they came from, so people can verify them
  • Retrieval respects the permissions of the person asking
  • Irreversible actions wait for human approval
  • Outputs are validated, with explicit handling for timeouts and refusals
  • Usage is metered per user, so cost is predictable
  • Every step is logged, so behaviour can be reviewed

How we deliver

  1. 01

    Define the use case

    One real task, the data it needs and how success will be measured, written down before any build.

  2. 02

    Build retrieval first

    Ingestion, chunking, embedding and search, tested on real questions before adding an agent.

  3. 03

    Add reasoning and tools

    Agent steps, API calls and approval gates, each with tests and logging.

  4. 04

    Evaluate and operate

    Ongoing evaluation on real usage, with monitoring for quality, cost and failures.

Proof from our own work

We build and run products ourselves, so these are systems we operate, not slides.

How we work with you

You deal with the engineers who do the work, not an account layer. We agree scope and priorities in writing, show working software on a staging site every week, and take your feedback into the next increment. Changes are welcome, and each one is discussed in terms of effect on scope and schedule before we start it, so there are no surprises at the end.

Everything we build lives in your own accounts: the source repository, the cloud environment, the domains and the credentials. If you later want to bring the work in house or move to another team, you can, because nothing depends on us holding the keys. Read more about how we work and the engagement models we offer.

Common questions

What is RAG?

Retrieval-augmented generation: the system finds relevant passages from your own data and gives them to the language model, so answers are grounded in your content and can cite it.

Can agents take actions in our systems?

Yes, through your APIs, with permissions respected and a human approval step for anything irreversible.

Which models do you use?

We choose per use case, including Gemini, behind a stable internal interface so the model can change without rewriting the product.

How do you stop the model from making things up?

By grounding answers in retrieved passages, requiring citations, validating outputs and keeping a person in the loop where the stakes are high. No system removes the risk entirely, so we design for it.

Where does our data go?

We design the data flow with you: what is indexed, where it is stored, who can retrieve it and which model providers see it. Access rules from your source systems are kept intact.