Technology · Applied AI
AI where it earns its place.
Forecasting, classification, document extraction, inference over incomplete telemetry and agent access. Each one is specific, testable and held to a baseline. Everywhere else, engineering.
A deterministic envelope around every model.
We use AI where it is specific and true: solar resource and yield forecasting, crop disease classification, inference over incomplete telemetry. Every model output runs inside a deterministic envelope with a defined failure mode.
A forecast is published with its error against a persistence baseline, a classification is published with its confidence, and neither is allowed to write directly into a system of record.
What the envelope forbids matters as much as what it permits: a model may not invent a missing input, may not trigger a financial movement, and may not bypass the operator override that sits on every automated step.
Input
Deterministic envelope
Defined failure mode
Output
What a model is held to.
- Held to
- A forecast is a capability only once it beats persistence on real data.
- Published with
- Its error, its confidence, and the record it was derived from.
- Never
- A number with no physical basis, presented as though it had one.
A model may not
- Write to a system of record
- Move money
- Invent a missing input
Seven systems, each stated by status.
We name AI where it is specific and true, and we say when something is research.
-
Pest and disease risk scoring and crop suitability, with advisory written in Hausa for the farmer portal. Scanned farm maps are digitised and georeferenced with AI.
-
ExportFlow
DeployedReads shipping documents by OCR and extracts typed fields. Every field is either Extracted or Unavailable, and the rules that grade findings are deterministic and versioned.
-
ClearMind
DeployedIris, a planning assistant built on the Gemini API, helps with planning and is not used as a source of record. Outside agents reach the workspace through an MCP server with scoped OAuth tokens.
-
Meridian Books
BuiltAn MCP connector lets an assistant work with the books under the signed-in user's role. Posting rules stay in the database, where a model cannot bypass them.
-
Career services platform
DeployedCounsellor case queues with AI-assisted handoff. A person decides; the model prepares.
-
NMAS
DeployedAI use is disclosed in the delivery package, alongside the versioned methodology for each indicator.
-
MaizeGuardAI
ResearchA convolutional network that classifies maize leaf diseases, exported to TensorFlow Lite to run on a phone without a connection. A research model, not a field deployment.
An agent is a user with a narrower role.
Autonomous agents are useful when they can act, and dangerous when they can act on everything. We give them the same thing we give a member of staff: an identity, a scope, and an audit trail.
- Model Context Protocol servers expose a product's real operations to assistants such as Claude, behind OAuth discovery.
- Scoped tokens mean an assistant can read or update tasks without holding the whole account, and can be revoked without changing a password.
- The same rules as the web app apply to an agent's request: permissions, validation, period locks and approval steps.
- Consequential actions stay with people. Releasing stock or adjusting a balance needs a second person, whoever or whatever asked.
How AI should behave in essential services.
Engineering positions, not policy commentary.
Contact
Have a model that needs to survive production?
Tell us what it predicts, what it is allowed to change, and what happens when its input is missing. We will tell you where the envelope belongs.