-
-
Learner
-
Parent
-
Teacher
Inspiration
As a mother of three navigating Kenya’s Competency-Based Education (CBE) system, I was overwhelmed by school notices: Messages and emails listing items for practical lessons, project materials, and assessments. I regularly forgot things, from plastic bottles for science to art supplies, and faced last-minute panic. Many Kenyan parents, learners, and teachers share this burden.
What it does
Msingi AI is a multi-agent system built on the Jaseci stack in the Jac language. It serves four users:
Teacher Copilot drafts KICD-aligned lesson plans and four-tier KNEC assessment rubrics (Exceeding, Meeting, Approaching, Below Expectations). Learner Companion builds termly revision schedules and material checklists. Teacher Taratibu is a Socratic tutor that guides learners through hard concepts without giving answers away. Parent Liaison turns technical curriculum notices into plain English guidance, with key takeaways in Kiswahili.
How we built it
The KICD framework is modeled as a connected graph of EcosystemRoot, GradeNode, and CurriculumNode objects, with edges linking learning outcomes, competencies, and materials to five benchmark years: PP2, Grade 3, Grade 6, Grade 9, and Grade 12. Walkers orchestrate the agents through byLLM. The interface is served directly from the Jac runtime with Tailwind CSS and a calm academic palette.
Challenges we ran into
Fitting the CBE curriculum’s strand-level outcomes into clean graph nodes took careful schema design. Teachers’ shorthand (PCIs, CBA rubrics) had to be translated into parent-friendly language without losing instructions. Payloads and the UI also had to stay light enough for low-bandwidth mobile connections. During development, the startup status banner intermittently showed “Failed to fetch” in the dev preview. We traced it to browser-to-server connectivity, not the model key.
Accomplishments that we’re proud of
We built a working multi-agent prototype from scratch in about 24 hours. It covers the five benchmark grade levels, PP2 through Grade 12, as a queryable graph. The Parent Liaison distills dense teacher notices into guidance a parent can act on the same day. Retrieval over the curriculum graph lets agents answer from stored KICD content”
What we learned
Graph-based retrieval combined with autonomous walkers in Jac is a powerful pattern for domain-specific agents. Technology in education must do more than digitize textbooks. It has to support collaboration between parents, teachers, and learners, and reduce the burden on the people doing the work.
What’s next for Msingi AI
Test with real parents and teachers in Kenya, and measure the time saved on notice-reading. Add Kiswahili as a full interface language, not only in takeaways. Support offline or SMS delivery for parents with limited data access. Expand the graph to more subjects and assessment tools, and validate content with KICD-aligned educators. Harden deployment with a sealed production bundle so service status stays reliable outside dev mode.
Built With
- byllm
- css
- jaseci
- tailwind
- vite
