The agent identifies patterns. The teacher decides what they mean.
MwalimuLens explores how an AI assistant can organise evidence of learning across terms, describe uncertainty and prepare qualified observations for a teacher's review—without making decisions about a learner on its own.
How it works
1. Retrieve evidence
Collect competency-specific assessments, practical tasks and teacher observations from multiple terms using controlled Model Context Protocol (MCP) tools.
2. Describe a bounded pattern
Reference source identifiers and distinguish supportive observations, conflicting records and unresolved questions. A result may be a candidate or an abstention.
3. Leave decisions to teachers
Any candidate remains pending teacher review. The model cannot approve a candidate or update a learner profile on its own.
Human judgement is part of the architecture
The project separates model-generated interpretations from application-enforced controls. Validation checks source references and response structure, but structural correctness is not treated as proof of educational accuracy.
- Uses only fictional synthetic data in development and demonstrations.
- Retains supporting and counter-evidence and explicitly records uncertainty.
- Restricts the model's ability to perform teacher-only actions.
- Tests abstention and failure cases alongside candidate proposals.
Inspect the work
The source repository includes architecture notes, test suites, evaluation documentation and development branches. The ongoing work focuses on reliability, source-grounding and the separation between model suggestions and human authority.