Skip to content
Syntrigen

Service 02

AI in production

An agent works in a notebook and fails in the world. No retry path, no permissions model, no record of what it did, and no way to tell whether last week's prompt change broke it.

Start a projectTypically 4 to 10 weeks

How we approach it

We build the agent as a production service: scoped tool access, explicit escalation, tracing on every run, and an evaluation suite that gates every change. This is the part of an AI-native product that decides whether it survives contact with real users.

What you receive

A deployed agent with tool permissions, retry and escalation behaviour, per-run traces, and an eval suite wired into your CI.

A concrete example

A support agent that reads the ticket, pulls order history, drafts the reply, and hands anything touching a refund to a person with its reasoning attached.

Related work

What this looks like when it is finished

All case studies

Our own product · AI in production

Live transcription with speaker attribution, running on our own infrastructure and our own pager

Mavio transcribes meetings in real time across desktop, mobile and a browser extension. The hard parts were never the model — they were system audio capture on two operating systems, speaker attribution over a noisy channel, and retention rules that differ per artefact.

TypeScript · Python / FastAPI · Electron · Expo · Supabase Postgres · Deepgram Nova-2 · AWS ECS Fargate · Terraform

How it runs

Four stages, in this order

The risky work happens second, not last, so a bad answer arrives while it is still cheap.

  1. 01

    Shape

    A written scope: what we would build, what we would not, and what could stop it.

  2. 02

    Prototype

    A working prototype on your data, and an honest read on feasibility.

  3. 03

    Production

    The product running in your infrastructure, with its evaluation suite and monitoring.

  4. 04

    Operate or hand over

    An agreed monthly scope, or documentation and access.

Before you write

We turn work down

It is cheaper for both of us to find out here.

A good fit

  • Your work has a shape the software you buy does not have, and the difference is costing real hours.
  • You want to own the product at the end of it, not rent a seat in someone else's.
  • Someone technical on your side can make decisions without a committee.
  • You would rather hear that a step is not feasible in week two than in month five.

Not a fit

  • You want a proof of concept whose purpose is to exist rather than to run.
  • What you actually need is an off-the-shelf tool, configured well. We will tell you when that is the answer.
  • You want a strategy deck with no build attached.
  • You need a fixed scope signed before anyone has looked at your data.

Start here

Describe the process and we will tell you whether this is the line it needs.

If it turns out to be none of them, we will say that too, and point you at whoever does it properly.

Start a projectOr start with a Production Readiness Review