Product engineering
Web applications people pay for: accounts, billing, permissions, dashboards — the unglamorous parts that decide whether a product survives its first hundred customers.
AI-first software studio · Newcastle, Australia
Arnavia designs and ships software with AI inside it — retrieval over your own documents, agents that act on real systems, classifiers that retire a queue of manual decisions. We also build the applications, integrations and data those systems depend on, because an AI feature with a broken billing path is not a feature.
01 — The gap
Not because the model is weak. Because the three things that decide whether it survives production were never built.
A language model is a component, not a product. It is non-deterministic, it fails in ways that look like success, and its cost scales with use rather than with seats. Design around those three facts and you get a system. Ignore them and you get a demo that gets quietly switched off six weeks after launch.
We have watched this happen from the inside of the integration. The retrieval “works” until the documents change. The assistant is impressive until someone asks it about a policy that was updated last Tuesday. The classifier is 90% accurate, which is excellent until you learn the 10% is the queue that matters.
The fix is not a better prompt. It is an evaluation set built from your own history, a deterministic path kept for the cases that must never be guessed, a spend ceiling per tenant, and a log of what the model was actually given.
02 — What you get
01
Right-sized, not enterprise-bloated. When the platform is missing the one thing you need, we scope it, price it and build it — in writing, before the first hour.
02
Findings are read at the source: the file, the line, and the path that reaches it. Hand it to your engineer and they can find it in minutes instead of arguing with it.
03
Your code, your data, your accounts, full access. Leave whenever you like and take everything with you — that is the entire argument for hiring a studio this size.
03 — Capabilities
We take the work where the deliverable is software in production, not a recommendation. If the honest answer is a smaller job, we quote the smaller job.
Web applications people pay for: accounts, billing, permissions, dashboards — the unglamorous parts that decide whether a product survives its first hundred customers.
Retrieval, assistants, agents and classifiers built with an evaluation set and a deploy gate, so “it works” is a measurement rather than an opinion.
APIs, webhooks, sync, single sign-on and migration between systems — including the ones that should have been retired in 2011.
Code and app review, QA as a continuous layer, and a three-week audit of how work actually flows through the business.
Pipelines that surface what the operation already knows — margin by client, order velocity, anomalies before they become incidents.
Architecture, AI readiness and a fractional CTO retainer for companies that have engineers but nobody above them making the technical call.
04 — How we work
We read what you run before we propose anything: the repository, the schema, the billing account, the queue. A fixed price follows from evidence.
Scope, data flow, risk register and the decisions that are expensive to reverse — written down, with the trade-offs named rather than implied.
Two-week sprints against a signed scope. Tests and evaluation run on every change; a regression fails the deploy rather than the customer.
Monitoring, spend caps, an audit trail and a documented path for the cases where the model must not be trusted. Support is a retainer, not a hope.
05 — What we will not do
A studio that says yes to everything cannot hold a fixed price. These four are the ones we turn down, and saying so is how the other fourteen stay honest.
Next step
No pitch deck, no mandatory demo. If we are not the right fit, we say so and suggest who is.
Fixed price, fixed date, agreed in writing before the first hour.