An 8-agent Claude system that personalises S1–S2 mathematics practice for Hong Kong schools — guiding students through sessions, helping teachers understand their class, and giving principals readable term reports.
S1–S2 students complete adaptive 20-minute sessions. Questions are chosen to target mastery gaps. Every interaction — correctness, time, hints, self-ratings — is logged.
8 specialist Claude agents process each event in real time: explaining questions, generating session summaries, updating gamification, and producing teacher insights.
Teachers see live class heatmaps and AI-written student profiles. Principals receive readable term reports — no data expertise needed.
Explains concepts and reflects on session performance with each student.
Summarises class mastery data and generates individual student learning profiles.
Cleans past-paper questions, estimates difficulty, and tags learning objectives.
Generates all in-app copy: welcome screens, session intros, question feedback, and gamification prompts.
Writes teacher onboarding guides, step-by-step feature walkthroughs, and FAQ answers.
Produces Mermaid diagrams — high-level architecture, student session flows, teacher upload sequences.
Generates badge messages, mission suggestions, and progress nudges to sustain motivation.
Writes clear, non-technical term reports for principals, panel heads, and school boards.
A single AdaptiveAgent orchestrator lazily instantiates each specialist, routes by mode, and resolves legacy aliases. All agents share a claude-haiku-4-5-20251001 model layer and return strict JSON schemas — no free-form text leaks into the application layer. Postgres with pgvector stores mastery state, session history, and question embeddings.