AI Agent Development
We build AI agents and machine learning systems that do a specific job inside your business: answering customers from your own documents, reading invoices, flagging a fault before a machine stops.
Most AI projects we are asked to look at have the same problem. The demo worked, and nothing after it did. A language model that answers well in a notebook still needs access to your data, guardrails around what it may do, a way to measure whether it is right, and somebody watching it once real users arrive. That part is ordinary software engineering, and it is where most of the effort goes.
So we start from the task, not the model. If a rules engine or a plain search index solves it, we will say so, because it will be cheaper to run and easier to trust. Where a model does earn its place, we pick the smallest one that meets the accuracy target, since model size drives both the monthly bill and the response time your users feel.
What we build
Task agents
Agents that work through a multi-step job using your systems as tools: look up the order, check the policy, draft the reply, hand over to a person when confidence is low. Every tool call is logged so you can see why the agent did what it did.
Retrieval over your own content
Question answering grounded in your manuals, contracts or knowledge base, with the source passage shown next to each answer. We spend more time on chunking, metadata and evaluation sets here than on prompts, because that is what decides answer quality.
Custom machine learning models
Classification, forecasting and recommendation models trained on your data with Scikit-learn or TensorFlow, for cases where a general language model is the wrong tool or too expensive per prediction.
Computer vision and EdgeAI
Models that run on the device instead of in the cloud. Our facial recognition system runs TensorFlow Lite and OpenCV on NVIDIA Jetson hardware, which keeps video on site and removes the round trip to a server.
How a project runs
- 1
Scoping
One or two weeks to agree the task, the success metric and the data we are allowed to use. You get a written scope before any build work is billed.
- 2
Evaluation set first
Before building, we collect real examples with known good answers. It is tedious. It is also the only way to tell later whether a change made things better or worse.
- 3
Prototype on real data
A working slice against your actual documents or sensor feeds, not sample data, so the awkward cases show up early.
- 4
Production build
Authentication, logging, cost limits, fallbacks for when the model provider is down, and a human handover path.
- 5
Handover
Documentation, runbooks and a recorded walkthrough for your team, as set out in our knowledge transfer guarantee.
The people you meet at the start are the people who build and hand over the project. That is the first of our six written guarantees.
Technologies we use
- Python
- TensorFlow
- TensorFlow Lite
- Scikit-learn
- OpenCV
- Google Cloud AI
- NVIDIA Jetson
Related case studies
Questions we get asked
How much does it cost to build an AI agent?
It depends mostly on how many systems the agent has to talk to and how accurate it must be before you can let it act without review. A retrieval assistant over one document set is a small project. An agent that writes to your ERP is not. Our AI Implementation Cost Calculator gives a first range in a few minutes, and a scoping call turns that into a fixed quote.
Do you train models from scratch or use existing ones?
Usually existing ones. For language tasks we integrate hosted or open-weight models and put the work into retrieval, tools and evaluation. We train custom models when the data is tabular, visual or time series, or when the model has to run on a small device.
Will our data be used to train someone else's model?
Not by us, and we configure provider settings so it is not retained for training on their side either. If your data cannot leave your network at all, we deploy open-weight models on your own infrastructure or on edge hardware.
What happens after launch?
Models drift as your products, prices and documents change. We set up monitoring against the evaluation set and agree a maintenance plan, or we hand the whole thing to your team with documentation. Both are normal.
Talk it through with an engineer
Thirty minutes, free, with someone who builds these systems. Bring the idea, the half-finished project or the vendor quote you are unsure about.