Book a call

Research

We test models before we trust them.

Our R&D team trains and evaluates models for work general AI handles badly: low-resource languages, private data, and decisions that have to come out the same way every time.

What we work on

Five areas, all tied to client work.

Speech and language

General speech models handle Urdu badly, and phone audio makes it worse. We work on Urdu speech recognition and voice, and the calling agent on this site is where that work runs first.

Talk to the voice agent

Computer vision

Proof of concept

Most businesses already have cameras that nobody watches. Our vision proof of concept turns existing camera feeds into signals a team can act on: activity, objects and safety events, without new hardware.

See Vision AI

Retrieval (RAG)

An assistant is only as good as the passages it finds. We test mixed retrieval approaches, combining different ways of finding and ranking sources, and keep the one that cites the right document most often on real questions.

Evaluation

A model change should be a measured decision. We build fixed test sets from real cases, score answers blind and run the same questions again to see how stable the results are.

Private deployment

Some data can’t go to a public AI service. We serve open-weight models on a client’s own hardware or private cloud, sized to their volume and tested before go-live.

Enterprise AI

How we evaluate

The rules we hold every model to.

  1. 01

    Real cases, frozen

    Test sets come from the work the model will actually do, and they’re frozen before a run, so two runs can be compared.

  2. 02

    Scored blind

    Model, run and setting labels are stripped before anyone reads an answer, so the scorer can’t favour the version they expect to win.

  3. 03

    Per measure, never one number

    Answers are scored on separate measures such as sources, structure and safety. A single headline score hides exactly the failures that matter.

  4. 04

    Some failures fail outright

    An unsafe answer or an invented source fails the answer, however well it scores on everything else.

  5. 05

    Repeat runs

    The same question in different phrasings, asked more than once. A model that’s right once and different next time isn’t ready.

  6. 06

    A manifest for every run

    Model version, settings and code version are recorded with each run. When a model is retired, the baseline says exactly what it measured.

Testing the system

Load-test the part that breaks.

Under load, AI products usually fail before the model is even called: in retrieval, in the processes they start, in memory. So we test the system with the model stubbed out, then measure real cost and latency on a small live sample. It finds the bottleneck without paying for hundreds of model calls.

Have a task or a dataset that general AI handles badly?

Send an enquiry