Home Services AI Product Engineering

AI Product Engineering

AI Product Engineering is the work of putting a model inside a product people use every day, and keeping it right after launch.

Overview

A product where the model does real work: it answers, drafts, reads, decides or acts. We build the model layer, the product around it, and the review a person needs before the output goes out.

Most AI pilots stop at a demo. We plan the production system from the first week: the data it needs, the people who review it, the cost it can carry and the way it is supervised after go-live. The first release is small and dated, and it goes to real users with an evaluation set behind it.

Who it is for

Enterprises with an AI pilot that has to become a product people rely on.
Product companies adding AI to a product customers already pay for.
Founders building a new product where the model is the core of the value.
Teams that have tried a chatbot and need something that does the actual job.

What you get

A ranked list of use cases, with the value, the risk and the data each one needs.
A target architecture: models, retrieval, tools, review and hosting.
Assistants and agents inside your product, not beside it.
Retrieval over your own documents and data, with sources shown to the user.
An evaluation set built from your real cases that says when the model is good enough to ship.
Guardrails, audit trails and a person who can take over.
Cost and latency budgets agreed before launch, and dashboards that show them after.
Documentation and handover, or a managed run after go-live.

How it runs

Discovery, two weeks: use cases ranked, data checked, architecture and cost model written down.
Evaluation first: we collect real cases and agree what a good answer looks like before building.
First increment, six to ten weeks: one use case, end to end, in front of real users.
Hardening: guardrails, monitoring, security review and load testing.
Go-live with a named reviewer and a rollback plan.
Run: we supervise, evaluate every change and widen the scope release by release.

In practice

A support assistant that drafts replies from the help centre and past tickets; an agent approves before sending.
A document workflow that reads claims or invoices, extracts the fields and flags what needs a person.
A sales copilot that prepares account briefs from the CRM, emails and call notes.
A B2B lending product where AI reads bank statements and a risk officer approves the credit decision.

Technology

Hosted models from Anthropic, OpenAI and Google, and open models such as Llama and Mistral.
Retrieval with vector search on Postgres (pgvector), Elasticsearch or a managed vector store.
Orchestration with LangChain, LangGraph or plain code where that is simpler.
Training and fine-tuning with PyTorch when a task should run on your own model.
Hosting on AWS, Azure or Google Cloud, in your account.

How we charge

A fixed-fee discovery, then fixed-scope increments with dates. After go-live, a monthly managed run or a handover to your team.

Questions

Do you work with our existing product team?

Yes. We can build the AI layer alongside your engineers, or own the whole product and hand it over.

Which models do you use?

Hosted and open models, chosen per task on your own evaluation set. The choice is written down, and the product can move to a better or cheaper model later without a rebuild.

Where does our data go?

It stays in your cloud account where possible. Hosted models are used under terms that do not train on your data, and that is checked before anything is sent.

What if the model is not good enough?

The evaluation set tells us early. If a use case cannot reach the bar, we say so and move to the next one on the list instead of shipping it.

Talk to the team

Two-week, fixed-fee discovery. Production release by week 12.

Book a discovery call