AI Operations — Colombo · Melbourne

Your operations, automated. You keep the controls.

Six to sixteen weeks, set by the process. We map it, agree the bar on your own cases, bound what agents can touch, and prove it before it goes live.

Send us a process Three engagements, Q4
Premise

An agent that passes a demo and a process that survives an audit are different problems.

The second one needs the work around the model: the contracts it runs under, the traces it leaves, the point where it stops and asks a person, and the number it has to hit on your own cases before any of it counts.

Ordinary software either works or it doesn't. Put a model in the middle and the same input can come back different — so we don't promise it works. We agree the pass rate, on work you've already done, before anything is built.

Practice

One engagement. The length is set by the bar.

If the steps are known before we start, we build a fixed path — cheaper, and far easier to hold to a number. We only build an agent that decides its own route where the route honestly can't be drawn in advance.

Workflow

Steps known upfront

A fixed path, with judgment at the decision points. Pipeline or graph. Most work should be this — and reaching for a harness where a workflow would do is how these projects run over budget.

Harness

Steps found at runtime

An agent in a loop with tools, deciding its own next step. For the questions where the path can't be drawn in advance. More capable, and considerably more expensive to gate.

The engagement
  1. 01

    Map

    • Every step, branch and handoff traced
    • Output Map and volumes
    • Golden set: your past cases, with what actually happened to each
    • The bar: a pass rate agreed with you, measured on your own past cases, before a line is written
  2. 02

    Bound

    • Deterministic plumbing moves the data. The model only makes the calls that need judgment
    • Tool contracts: typed schemas, error states, success criteria
    • Deny by default. Traces on every run
    • When it isn't sure, it stops and asks a person. That threshold is yours to set
  3. 03

    Prove

    • Run against the bar until it clears — then two weeks in shadow mode on live work
    • It suggests, your people confirm, nothing it says is binding
    • The same tests keep running against live traffic after launch
    • Working in March, not just in the demo
Terms, stated up front

Priced before built. Your account, your bill, your bar.

Built in your cloud account from day one. Your data never leaves your tenancy. The handover is the infrastructure, the dashboards, the runbooks and the test suite. Your IT owns it.

Running costs are billed to you directly, at cost — cloud, compute, tokens, all of it on your own account. We don't resell infrastructure or mark up usage. Expect tens of dollars a month at a thousand items a day. No idle GPU bills, no per-seat licences.

It watches itself and ships its own routine fixes. Nothing here needs daily attention. What it does need is a person a few days a quarter, and an owner for the day a model is retired. We'll train your team to do that, or do it ourselves.

Records
“Every organisation is carrying a problem nobody has looked at end to end. The work is to understand it fully, then find the smallest change that moves it.”
Tanisha Wijesinghe

Tanisha Wijesinghe

Fintech experience

Engineer-turned-designer. Tabcorp, Yondr Money, Harvard FinTech. She maps the process before an agent is allowed near it.

Dr. Sithira Ambepitiya

Dr. Sithira Ambepitiya

Agent systems

Pass rates, traces, evals. The part of the system that has to survive an audit, not just a demo.

Enquire

Tell us the process.
We will tell you what it is worth.

Send us a process