
Manufacturing
Vision inspection that catches defects a tired operator will miss, and maintenance scheduled from sensor drift instead of a calendar.
AI product studio — independent
Four products live in the world — a Shopify SaaS and three mobile apps — each with AI running in production, not in a demo. That is the same team you get for your build.
We operate our own products. That means we carry the support load, the inference bill, the App Store review, the churn when a feature misses. Every lesson on this site was paid for.
It shows up in the choices we make on client work. Agents get approval gates because we learned what an ungated agent can do to a live storefront. Results get holdout groups because we know how easy it is to claim a lift that was never there. We would rather report a smaller number you can trust.
We take on a small number of client builds a year, alongside our own. That is the constraint that keeps both good.
Not case studies from a client list — products we own, operate and answer for. Every one of them is live right now.
A multi-company ERP for a T-shirt manufacturer: roll-level fabric stock, production with job-workers, GST sales and purchase, receivables and payables — built from the owner’s process, not a template.
We would rather be excellent at a short list than available for everything.
Not the demo on stage. The line that stops at 3am, the reconciliation nobody wants to own, the forecast somebody still does in a spreadsheet.

Vision inspection that catches defects a tired operator will miss, and maintenance scheduled from sensor drift instead of a calendar.

Demand forecasting, slotting decisions and exception handling — the three places a warehouse quietly loses hours every single day.

Getting SCADA and PLC data somewhere a model can actually reach it, then flagging anomalies early enough to matter.

Order-to-cash, reconciliation and reporting across systems that disagree with one another. Usually the fastest payback of anything here.

Agents that sequence and supervise a cell, with a human holding the gate on anything that moves a physical thing.

Where the models actually run. On-premise or in your own VPC when the data cannot leave, with the cost per inference measured.
Photography is illustrative — sectors we build for, not client sites
Talk about yoursNo six-month discovery. You should know whether this is working within a fortnight.
We get access, instrument what is not measured, and come back with the specific thing that is costing you — not a list of everything that could be improved.
The narrowest build that proves the value is real. It runs on your data, and it is allowed to fail here, where failing is cheap.
Production hardening: evaluation suites, monitoring, cost controls, approval gates and rollback. This is the part most AI pilots never reach.
We run it, or we hand it to your team documented and legible. Both are fine. Being quietly indispensable is not the goal.
Not a discovery call about your digital transformation journey. A real problem, and a straight answer about whether we can fix it — including when the answer is no.