Taking two new engagements this quarter

Your business, run on AI

We find the work AI can take off your team's plate, build it into how you already operate, and keep it running. Senior engineers from the first call to go-live.

Book a demoSee how it works

30 minutes with the engineers, not a salesperson. We'll tell you if it's not a fit.

ai-division://runtimeIn production
01

Ingest

Your data, your systems

02

Reason

Models under evaluation

03

Act

Into the tools you use

GuardrailsEvalsObservabilityHuman-in-the-loop

Running in production for

  • StudentQR
  • PBLSH
  • Braintree Technologies
  • The Mantra
  • Alomaid
  • MagNicas
  • REKA
What we build

AI that does real work, not demos

Six kinds of system we put into businesses. Every one of them is running today, for a client or for ourselves.

Assistants that answer your customers

On WhatsApp, day or night. They answer what they can and hand the rest to a person, with the whole conversation attached.

Zahra, for StudentQR

AI that reads your paperwork

Receipts, bank statements, even video. It pulls out what matters and fills in the forms nobody wants to type.

Books and PBLSH

Ask your business a question

Connect Claude or ChatGPT to your own data over MCP, and ask how sales went this month in plain words.

The Mantra

Automations that run themselves

Updates, reminders and hand-offs between your tools, each drawn as a diagram, with an alert the moment a step fails.

Automator

The software around the AI

Dashboards, approvals and internal tools, so the AI lands inside an app your team already opens every day.

The Mantra and SecureTrace

A person signs off

The AI drafts, someone approves. Anything that touches money, customers or records waits for a human yes.

Books and PBLSH

Selected work

Already live. Already paying off.

All case studies
AI SquadIn the works

Meet the AI Squad. Specialists, not a chatbot.

We're building a team of AI agents where each one is great at a single job. A lead agent breaks your request down and hands every piece to the specialist who knows it best. Nothing leaves the squad without a person's OK.

Lead

Splits the job, hands it out

  1. Support

    Answers customers on WhatsApp

    Working

    Answer a customer asking where their order is

  2. Analyst

    Answers questions from your data

    Working

    Pull this month's sales by channel

  3. Bookkeeper

    Drafts entries, matches the bank

    Working

    Match 42 bank lines to the books

  4. Ops

    Runs workflows, fixes failures

    Working

    Re-run the delivery sync that failed at 3am

  5. Reviewer

    Checks every result against evals

    Working

    Check every draft before a person sees it

  6. Content

    Writes titles, tags and listings

    Working

    Tag the 18 video clips uploaded overnight

  7. You

    3 drafts ready for your OK

A preview of what we're building. The tasks are examples.

  • 01

    One job each

    Every agent has its own instructions, tools and limits, like a new hire with a clear role.

  • 02

    A lead that delegates

    It splits the work, routes each task, and brings the results back together.

  • 03

    Grown from what already runs

    Each specialist builds on a system we already have in production.

Under the hood

Safe enough to run your business on

Anyone can wire up a model. What makes AI safe to hand real work is everything we build around it.

  • Measured, not guessed

    An eval suite runs on every change, so we know a tweak made things better before it ships.

  • Only the tools the job needs

    An assistant that can request stock can't move stock or take money. Limits are part of the design.

  • People approve what matters

    Drafts wait for a person. Nothing is posted, sent or paid on the AI's word alone.

  • Every step on the record

    Each run is logged step by step, with an alert the moment one fails.

  • Not tied to one model

    Claude, GPT, Gemini: we pick per task, and switch when a better one arrives.

  • Inside your walls

    It runs within your systems and security boundary, and your data isn't used to train anyone's model.

books · mcp traceReplay
  1. telegramreceipt.jpg

    A receipt arrives on Telegram

  2. → list_files()1 new

    Finds the new receipt

  3. → get_file()

    Reads it: vendor, date, total, tax

    vendor
    Bumi Coffee Roasters
    date
    2026-09-19
    ref
    BR-2291
    total
    RM 535.30
  4. → list_accounts()Cost of goods sold

    Picks the expense account

  5. → create_draft()status: draft

    Saves a draft entry, not posted

  6. ownerpost

    The owner checks it and posts

Posted by a person6 steps · 0 posted by AI

A replay of one receipt through Books, with its real tool names. The receipt itself is an example.

How it works

One workflow at a time. Live in weeks, not years.

No two-year roadmap. We pick one workflow that matters, get it live, then take the next, with the same team the whole way.

01

Find

One week mapping where AI pays off in your operations, and saying plainly where it doesn't.

02

Build

A working system in your stack, in front of real users, within six weeks.

03

Spread

We stay on, keep the first one sharp, and take on the next workflow.

Why us

Most AI transformations stall right after the pilot

The demo lands, everyone's impressed, and then nothing reaches the business. We're built for the hard part: getting AI into daily operations and keeping it there.

The people who pitch it are the people who build it

No account layer, no hand-off to juniors, no one learning on your budget. The senior engineers on your first call are the ones writing the code six weeks later.

Runs on what you already have

We build inside your systems and your security boundary. Nothing to migrate, no new vendor to manage.

You'll know whether it's working

Every change ships behind an eval suite, so improvement is measured rather than claimed.

It spreads instead of stalling

Models drift, workflows change, and real usage never matches the pilot. We stay on to keep the first system earning while the next one goes live — which is the whole difference between a transformation and a pilot.

Questions

What people ask on the first call

Do we need clean data before we start?

No. Part of the first week is finding out what data you have and whether it's enough. If it isn't, we'll tell you what it would take before you spend anything on building.

Which AI models do you use?

Whichever does the job best for the cost: Claude, GPT, Gemini, or an open model you run yourself. Systems are built so the model can be swapped without rebuilding everything around it.

Will our data be used to train AI models?

No. We use the providers' business terms, under which your data isn't used for training, and keep the system inside your own infrastructure wherever you need it there.

What happens when the AI gets something wrong?

It's designed for that. Anything that matters waits for a person to approve, every step is logged, and a failed step raises an alert. Evals catch regressions before a change reaches you.

Will it work with the tools we already use?

That's the point. WhatsApp, Telegram, your accounting system, spreadsheets, internal databases: we build into what your team uses rather than asking anyone to move.

How long until something is live?

On most first projects, a working system in front of real users within six weeks. Bigger ones are split up so something useful ships early.

What does it cost?

It depends on the workflow, so we don't quote blind. After the first call you get a clear scope and price for the first phase, and an honest answer if it isn't worth doing.

Tell us where the business hurts. We'll tell you if AI fixes it.

Thirty minutes with the engineers who'd actually build it. You leave with an honest read on whether it's worth doing, what it would take, and what it would cost.

Book a demo