Services · 04 - Persona & Knowledge Engineering
Most organisations doing serious AI work have the same problem. There’s no version. No evaluation. No governance metadata. No way to say what was live last Tuesday.
We replace that with a discipline.
It’s for organisations with deep, specific know-how - how this credit decision is made; how this bid is qualified; how this claim is handled - that they want encoded reliably rather than re-prompted ad hoc. And for organisations running more than one AI assistant or agent that need them consistent, governed and version-controlled instead of scattered.
Persona Factory: the engine underneath →What you get
A library of governed AI personas, served from Persona Factory, the Katemba platform component that authors, versions, compiles, evaluates and serves AI personas, skills, domain packs and output contracts.
The role, the boundaries, the tone, the decision logic, the outputs it can produce.
Representative test cases with expected behaviour, run on every persona version before it’s promoted.
What’s allowed, what isn’t, what data the persona can see, what tools it can call, who can change it.
Every change, who made it, when, why.
Served over the Model Context Protocol, so any compliant agent or assistant can consume the persona without custom integration.
The platform component underneath this service: authoring, compilation, evaluation and serving, and how they fit together.
Read on the platform page →The difference
“We pasted a prompt into a chatbot.”
“Here is the tested, version-controlled model of how this decision gets made, and the evidence it behaves.”
What we actually do
We sit with the people whose knowledge we’re encoding. We watch the work get done. We extract the implicit rules, the edge cases, the escalation triggers, the tone, the “we never do X” lines. This is domain analysis as much as it is prompt engineering.
We translate the expertise into governed personas in Persona Factory: explicit behaviours, explicit constraints, explicit tools and data access. Each persona is a small, focused unit; we resist the temptation to build one giant persona that does everything.
For every persona, we build a representative test set with the expertise owners: real cases, expected behaviour, edge cases that should be refused. This is the spec the persona will be measured against forever.
Personas are compiled and versioned. Each version is tested against the evaluation set before it’s promoted. Production traffic only sees versions that have passed.
Once personas are live, we monitor their behaviour, watch for drift, add cases to the evaluation set as the work evolves, and re-test against every model upgrade.
Why this matters
Different teams write different prompts for similar tasks. The behaviour of “the AI assistant” is whatever each team’s last prompt edit was.
Frontier models change every few months. Without evaluation sets, you find out the assistant got worse only when a user complains - and you can’t tell which change caused it.
When the regulator or the auditor asks what the system was doing on 14 March, the answer should be a version number and an evaluation report, not “we’d have to ask whoever was on the team that month”.
How we engage
We work with one team to encode one or two flagship personas - the ones that anchor the rest. End state: governed, evaluated personas live, with the engineering pattern your team can extend.
For organisations with a portfolio of AI work to bring under governance, we build out the persona library, typically 8 to 20 governed personas, with shared knowledge packs, evaluations and operating patterns.
For internal teams that want to own this themselves, we deploy Persona Factory and enable your engineers to build, evaluate and serve personas using it.
See Platform Enablement →If you can describe how the work gets done - and what you’d never let the AI do - the first workshop is already half-planned.