Applied AI that makes a decision
measured
Most AI work fails at the same place: it produces something impressive in a meeting and nothing an organisation can act on by Tuesday. We build the second kind.
Start where the signal already is
Organisations produce far more usable signal than they read — printed tickets, scanned forms, inbox threads, years of operational records. The first question is never which model to use; it is which decision is being made badly today, by whom, and what they would need in front of them to make it well. That answer usually names a narrow, measurable task, and narrow tasks are the ones that survive contact with production.
Accuracy is a number or it is an opinion
Every model we put into an operation ships with an evaluation harness: a held-out set, a score, and a threshold below which the system defers to a person instead of guessing. That is what makes it possible to change a prompt, a model or a pipeline later and know within minutes whether you made it better or worse. Without it, a system that quietly degrades looks exactly like one that works.
A person stays in the loop where it matters
An extraction that feeds a payment, a classification that closes a case, a summary that becomes part of a record — these need a route for a human to see, correct and approve, and an audit trail of who did. We design that route first. It is also what makes adoption possible: staff trust a system that asks before it acts, and stop trusting one that is confidently wrong once.
- Document and ticket OCR pipelines
- Forecasting, anomaly and fraud detection
- LLM assistants grounded in your own records
- Evaluation harnesses, so accuracy is a number
- 01A scoped decision, with the current cost of getting it wrong written down
- 02A pipeline in production, with the thresholds that send a case to a person
- 03An evaluation harness and a baseline score you can re-run
- 04A runbook for retraining, re-scoring and rolling back
Someone is retyping what is already printed on a document
A decision depends on a pattern nobody has time to look for
You have been shown an AI demonstration you could not put into the operation
- 02
Consignment Workflow Platform
An AI-assisted platform that replaced paper weighing tickets and manual release approvals end to end — web console, API and field mobile app.
- 08
Institutional Records Platform
A public platform and a full records system behind it, covering five distinct staff and student roles on one data model.
Do you train your own models or use an existing one?
Whichever the task needs. Document extraction is usually a pipeline around an existing model; forecasting and anomaly detection on your own operational history are usually trained. The decision follows the measurement, not the fashion, and we say which one we are proposing and why.
Does our data leave our systems?
Only if you decide it does, and we design for it not to by default. Where a hosted model is the right tool, we tell you exactly what is sent, what is retained and what the alternative costs. Where it is not, the pipeline runs on infrastructure you control.
What happens when the model is wrong?
The system says so and gives the case to a person. A confidence threshold, a review queue and an audit trail are part of the build, not a later addition — a system that cannot be wrong safely has no business touching a record.
How do we know it is still working in six months?
The same harness that scored it at launch runs against new data, and the score is visible to you rather than to us. Silent degradation is the characteristic failure of these systems and the only defence is measurement you own.
Start here
Bring us the problem nobody has solved.
A 30-minute technical briefing. We will tell you what we would build, roughly what it costs, and honestly whether we are the right people for it.
- ask@madsoftech.com
- Telephone
- +220 872841578
- Office
- Brusubi, The Gambia