Skip to content

/03

Applied Intelligence

A model is a component, not a product. We use one where it genuinely beats a rule — triage, extraction, summarising a queue nobody has time to read — and we keep a deterministic path underneath it so the business does not stop when the model is wrong.

RetrievalEvaluationHuman in the loop
Typical timeline
3–8 weeks
Starts at
from R64 000

What is included

  • A scoped evaluation set before any prompt is written
  • Retrieval over your own documents, with citations back to the source
  • A confidence threshold that routes uncertain cases to a person
  • Cost and latency budgets enforced in code
  • Prompt and model versioning, so a regression is traceable
  • A plain-language write-up of what the system cannot do

How it runs

  1. Qualify

    We check whether the problem needs a model at all. A third of the time the honest answer is a lookup table, and we say so.

  2. Measure

    A labelled set drawn from your real cases. Nothing ships without a number to beat.

  3. Build

    Retrieval, prompts, thresholds and the human hand-off, wired into the tools your team already uses.

  4. Review

    Live accuracy tracked against the evaluation set, with a monthly review of the cases it escalated.

Work in this line

Start a project