Book a demo

CITRIS AI Engineering

Agents that work alongside what you already run.

Custom agents, frameworks and accelerators that sit beside your existing applications and databases. We build inside your chosen cloud boundary and evaluate each agent against a defined test set before release.

Talk to us about custom AI See it in Artec Neo →

CITRIS AI Engineering

We design and build custom agentic AI.

Custom agents, frameworks and accelerators that sit beside your existing applications and databases.

We build agents around the applications and databases you already use. The engineering discipline is the same one that runs inside Artec Neo: evaluation before release, audit trails, cost caps, and human approval on anything that writes.

  • Ingest

    Data onboarding agents

    Point them at a legacy database, spreadsheet estate or API. The agent infers a schema and proposes the mappings, and a person approves them before the deterministic import runs.

  • Extract

    Document-to-record extraction

    O&M manuals, contracts and certificates become structured, evidenced data, with per-field confidence and a citation back to the source page.

  • Decompose

    Obligation intelligence

    Contracts and SLAs decomposed into dated, owned, trackable obligations, so nothing lives only in a PDF nobody has opened since signature.

  • Query

    Ask-your-systems

    Natural-language answers over your own data. For read-only use cases, the agent is not given write tools.

  • Build

    Custom agent frameworks

    Designed around your workflows and deployed in your own cloud: AWS Bedrock, Azure AI Foundry, or your own endpoints. Data and model calls stay inside the environment you nominate.

Engineering controls we apply to every agent

  • Agents propose, people approve

    Agents draft and deterministic code executes, with a person approving anything that writes. The model is not the last step between an idea and your database.

  • Deployed in your own cloud

    Bring your own model on AWS or Azure. We build to your boundary rather than asking you to move inside ours.

  • Measured accuracy

    Agents ship with a measured accuracy number against a held-out set, and a regression gate that fails the build when it drops.

  • Governed by default

    Audit trails, cost caps, tool allow-lists and PII redaction. The control plane is part of the build, not a phase-two extra.

Artec Neo is where we prove this first. Evaluation gates, audit chains, human approval and cost governance are implemented in our own platform, and NeO is being evaluated there before wider release.

How we engage

A three-stage engagement, starting with a two-week feasibility sprint

Sometimes the most valuable outcome of the sprint is the finding that the idea does not work. We would rather establish that in a fortnight than over a year.

  1. 01

    Feasibility sprint

    2 weeks

    We take one real workflow and one real dataset, and establish whether an agent can do the job to a measurable standard. You get an accuracy number, a cost-per-run figure, and a straight recommendation. Sometimes that recommendation is not to build it.

    You get Working prototype, evaluation set, go/no-go with numbers

  2. 02

    Build and evaluate

    6–12 weeks

    The agent is built against your systems. Audit, cost caps, tool allow-lists and PII handling go in from the first commit rather than being retrofitted before go-live.

    You get Deployed agent in your cloud, regression gate in your CI

  3. 03

    Operate or hand over

    Your call

    We run it, or we teach your team to. Either way the evaluation harness stays, so drift shows up as a failing build rather than as a complaint from a user six months later.

    You get Runbooks, eval maintenance, or a managed service

Straight answers

Four claims we will not make

  • “The agent will handle it end to end.”

    Agents are excellent at proposing and terrible at being accountable. We put deterministic code between the model and anything that writes.

  • “We’ll fine-tune on your data.”

    Usually the expensive answer to a retrieval problem. We reach for fine-tuning last, and only when an evaluation shows retrieval genuinely cannot get there.

  • “It works — look at these examples.”

    We agree the evaluation set before development starts, and we report the number it produces, including when it disappoints.

  • “Send us your data and we’ll take care of it.”

    We build inside your own boundary, using your keys and your cloud. Data leaving your control should be a decision you made rather than a side effect of a vendor’s architecture.

Next step

Bring us one difficult workflow.

The repetitive one where somebody re-keys data between two systems, or reads a hundred PDFs to answer one question. That is where an agent earns its keep.

Book a demo Talk to us about custom AI

A working session with the people who built it.