Expertise
AI and machine learning
IncogX uses AI where it improves a job a person already has, and declines it where it would only add a confident wrong answer. Evaluation, consent, and a named owner come before a demo that impresses a meeting.
In brief, then in detail
- A job, then a model. Classification, search, drafting, or a forecast. The feature is accepted because a workflow got better, not because a model exists.
- Retrieval before invention. Where the answer must come from the organisation’s own material, retrieval is designed first. The source is shown. Hallucination is treated as a defect.
- Evaluation. A set of cases, including the awkward ones, checked before release and after the model or the prompt changes. Accuracy is measured, not felt.
- Consent and accountability. Training and prompts use data the organisation has a right to use. A person remains responsible for a decision that affects a customer, a patient, or an employee.
How it is run
An AI feature is a backlog item with acceptance criteria, a risk note, and a way to turn it off. It does not bypass security review because it is new. Bias and error are discussed in language a sponsor can follow: who is harmed if this is wrong, and how we would know. Portfolio language applies: it is a hypothesis, funded while the evidence holds.
First questionWhich job gets better?
EvidenceA test set, not a demo
ControlA person owns the decision
Where it is used
- LLM integration
- Retrieval
- LangChain
- Evaluation sets
- NLP
- TensorFlow
- PyTorch
Terms, in plain language
- Large language model. A model that generates language. Useful for drafts and search. Not a system of record, and not an authority.
- Retrieval. Finding the organisation’s own passages first, so the answer is grounded in material you can open.
- Evaluation set. Examples, with the answer you expect, used to see whether a change helped or quietly made things worse.
- Human accountability. A named person can accept, reject, or override the output when the consequence is serious.