Software architect, AI-native education platforms.
I’m Francis, a senior engineer with 7+ years in TypeScript, React, Python and PostgreSQL. I design agent-native systems, where an AI agent and the user interface are equal partners over one set of operations, and I’ve taken two education products from an empty repository to a working platform on my own: Ellai, an AI study platform, and a school operating system for African schools.
Project
Agent-Native School Platform
Run a school by asking.
School software usually means menus, buttons and long forms. I wanted to replace that with a conversation. A principal, teacher or learner says what they need and an AI agent does it: it sets the school up, builds the timetable, drafts lesson notes, writes and marks work, and issues report cards. The agent can be the school’s own assistant, or the person’s own Claude or ChatGPT connected over MCP.
It is built for Nigerian secondary schools and portable to any country’s structure. The code and content were written in sessions with Claude Code under my direction: I set the scope, reviewed the output and tested it.
- Role
- Principal Engineer, solo, Jan 2026 – present
- Stack
- TypeScript, React Router, Node.js, Drizzle, SQLite/PostgreSQL, MCP
- Scale
- 200+ actions, 3 portals, 4 roles
- Demos
- 2 schools, 180 learners and 30+ staff each, a full term taught
What I tried to do
- 01
One set of actions, three ways in
Every operation is defined once as an action (200+) and served three ways: the React screens, the in-app AI agent, and external agents such as Claude, ChatGPT, Claude Code or Cursor over the Model Context Protocol. Whatever one does, the others show at once.
- 02
Same person, same permissions
Authorization is a single guard over every action, with four school roles and row-level class scoping. A teacher cannot reach another teacher’s gradebook by screen, HTTP endpoint or agent tool call. The agent drafts first and confirms before anything irreversible.
- 03
Curriculum to report card
Subjects, units and week-by-week pacing; a syllabus import that keeps per-objective provenance; weekly lesson notes drafted from the agreed pacing; activities from flashcards to timed exams with difficulty-tiered papers; rubric marking, gradebook, report cards and per-objective analytics.
- 04
Any country, no fork
Terms or semesters, year-group names, grading scales, pass marks and exam weeks are school data. I proved it with two demo schools on the same code: Crestfield College (Ibadan, WAEC grading) and Yellow View High School (Columbus, GPA on a 4.0 scale). Moving to the US took three platform changes; the rest was data.
- 05
Tested by simulation
A whole school runs for a term while learners carry a hidden ability, trajectory and diligence the app is never told. Where the app’s verdict and the truth disagree, that is a finding. It caught lost parallel invitations, learners marked out of papers they did not sit, and a student able to read unpublished work.
Try it
Both demo schools are live with simulated data; no real learner or teacher appears. Email me for the sign-in details.
Work
- Company
- Agent-Native School Platform
- Role
- Principal Engineer
- Period
- 2026 — Present
- Company
- Elite Software Automation
- Role
- Senior Software Engineer
- Period
- 2023 — 2025
- Company
- Fulfil.io
- Role
- Senior Frontend Engineer & Tech Lead
- Period
- 2021 — 2022
- Company
- Opendeskng Ltd
- Role
- Lead Frontend Engineer
- Period
- 2020 — 2021
- Company
- HyperionDev
- Role
- Technical Interviewer & Code Reviewer
- Period
- 2020 — 2021
- Company
- Andela
- Role
- Software Engineer
- Period
- 2019