HYPER-PERSONALIZED · CROSS-SUBJECT · CROSS-GRADE FRAMEWORK
One curriculum. A different path for every learner.
LalaSystem is a hyper-personalized, cross-subject, and cross-grade adaptive learning framework that transforms structured curriculum content into a living learning experience for every individual.
Available daily from 20:00 China Standard Time (UTC+8).
Features · Experience · Architecture · Quick Start · Documentation · Citation
Important
Claim boundary. The public demo provides a Grade 9 mathematics reference experience. Cross-subject and K-to-Gaokao expansion is the framework's design scope; each new domain requires aligned content, a maintained knowledge graph, subject adapters, teaching policies, and validation. The showcase opens daily from 20:00 China Standard Time (UTC+8).
| 01 · Grounded AI question generation A multi-agent pipeline uses the target graph node, current mastery, bottleneck, desired difficulty, prior wrong answers, and existing content to create, review, and deduplicate learner-specific questions for curated content expansion. |
02 · Hyper-personalized question routing Every round can follow a different concept path, difficulty band, question sequence, and review rhythm according to the learner's live state. |
| 03 · In-question Socratic tutor The assistant sees the active problem and previous guidance, then delivers focused prompts, hints, and stepwise scaffolding around the learner's own reasoning. |
04 · Personalized multimodal RAG PhotoQA combines the uploaded problem with semantically retrieved curriculum context, the learner profile, and recent learning history to ground each response. |
| 05 · Handwritten-work assessment On supported problems, learners can upload handwritten solution steps. A multimodal evaluator returns structured grading feedback and records the result inside the learning loop. |
06 · Living knowledge graph A maintained curriculum graph is overlaid with mastery, stability, confidence, transfer, bottlenecks, and forgetting risk for every learner. |
| 07 · Generative interactive canvases A dedicated visualization agent turns problem context into purpose-built 2D or 3D experiences for functions, geometry, vectors, surfaces, and more. |
08 · Closed-loop memory and review Goals and interactions persist across sessions. Wrong answers return on mastery-aware schedules, while sustained recovery graduates them from the review loop. |
One front door connects adaptive practice, intelligent review, multimodal Q&A, learner insight, and the next best learning action.
TaForYou is the current mathematics reference experience powered by the LalaSystem framework.
The problem stays in focus on the left. On the right, a context-aware teaching assistant sees the current question and guides the learner through a Socratic dialogue with questions, prompts, and stepwise scaffolding that preserves the learner's own reasoning process.
The curriculum graph defines concepts, dependencies, and progression. LalaSystem adds a continuously evolving state layer for each student: mastery, stability, confidence, transfer, forgetting risk, bottlenecks, and next actions.
The graph directly routes teaching decisions.
PhotoQA turns the problem context into a purpose-built interactive canvas, so learners can see and explore functions, geometry, and mathematical relationships.
LalaSystem personalizes at the student × concept × moment level.
Knowledge graph
× current mastery and stability
× recent mistakes and hint behavior
× prerequisite gaps and forgetting risk
× learner goals, confidence, and transfer
─────────────────────────────────────────
= the next target · difficulty · question · teaching move
Two learners in the same class can move through different concepts, difficulty bands, question sequences, and review schedules. Every new interaction enriches the learner state that drives the next decision.
The platform combines personalized retrieval from vetted content with a grounded question-generation module. The generation contract can use the target graph node, learner mastery, bottleneck, desired difficulty, prior wrong-answer examples, and existing content to expand the right part of the curriculum.
New learning vertical
= Content Pack
+ Maintained Knowledge Graph
+ Teaching & Assessment Policies
+ Subject Prompts and Adapters
Reusable platform
= Learner Model · Adaptive Engine · Practice · Review · PhotoQA
+ Socratic Tutor · Visualization · Analytics · Identity · Persistence
Each subject and grade defines its own domain layer. Teams bring aligned content, concept relationships, policies, and prompts; LalaSystem supplies the shared closed-loop learning infrastructure.
LalaSystem is the platform. Subjects are extensions. The same core can grow into LalaEnglish, LalaChemistry, and other domain experiences while reusing the learner intelligence, teaching surfaces, and data foundation.
The current reference application demonstrates selected mathematics journeys. The framework is designed to expand across subjects and from foundational learning to Gaokao-level curricula as validated domain packs are added.
flowchart LR
KG[Curriculum<br/>Knowledge Graph] --> STATE[Living Learner<br/>State]
SIGNALS[Answers · Mistakes<br/>Hints · Speed] --> STATE
STATE --> DECIDE[Target · Difficulty<br/>Teaching Move]
DECIDE --> EXPERIENCE[Practice · Tutor<br/>Review · Canvas]
EXPERIENCE --> SIGNALS
The graph explains what can be learned. The learner state explains what this student needs. The loop turns both into the next meaningful interaction.
.
├── .github/workflows/ Automated repository checks
├── agent/ FastAPI, orchestration, learner state, and persistence
├── frontend/ React + Vite learning experience
├── knowledge_graph/ Curriculum concepts and relationship graph
├── qa/ Question bank, migrations, and generation workflow
├── student/ Multidimensional learner-model schema
├── platform/ Identity and platform-persistence schema
├── assets/readme/ Sanitized product captures
├── tests/ Offline regression and integration tests
├── tools/ Content, migration, and research utilities
├── scripts/ Repository checks
├── LICENSE Apache License 2.0
├── PROJECT_METADATA.yml Machine-readable project metadata
└── REPRODUCIBILITY.md Setup and reproducibility boundary
Only directories used by the current implementation are retained. Runtime databases, provider credentials, user uploads, caches, and build outputs remain outside version control.
Backend and adaptive core · Reproducibility · Security policy · Project metadata
Requirements: Python 3.12, Node.js 20, and npm.
git clone https://github.com/HKUST-FintechLab/LalaSystem.git
cd LalaSystem
python scripts/check_repository.pyExpected result: [repository-check] PASS. This first check uses no external
provider and requires no API key.
macOS / Linux:
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
(cd frontend && npm ci)Windows PowerShell:
python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -r requirements.txt
Push-Location .\frontend
npm.cmd ci
Pop-LocationCopy the safe templates, then replace every URL, API key, and model with values from your own compatible providers. Real configuration files stay outside Git and the Docker build context.
cp agent/config/coordinator_llm.json.example agent/config/coordinator_llm.json
cp agent/config/photo_qa_llm.json.example agent/config/photo_qa_llm.json
cp agent/config/photo_qa_embedding.json.example agent/config/photo_qa_embedding.jsonOCR is optional; copy ocr_config.json.example only when needed. Windows users
can perform the same operation with Copy-Item.
Start the backend and frontend in two terminals from the repository root.
macOS / Linux:
# Terminal 1 · API
source .venv/bin/activate
python -m uvicorn agent.api:app --host 0.0.0.0 --port 8000# Terminal 2 · Web
cd frontend
npm run devWindows PowerShell:
# Terminal 1 · API
.\.venv\Scripts\Activate.ps1
python -m uvicorn agent.api:app --host 0.0.0.0 --port 8000# Terminal 2 · Web
Set-Location .\frontend
npm.cmd run devOpen http://localhost:5173. The API runs at http://localhost:8000; Vite
forwards /api and /media requests to it during development.
- Knowledge-graph-native curriculum architecture
- Dynamic concept-level learner state for every student
- Personalized planning, question retrieval, and review scheduling
- Grounded personalized question-generation module
- In-question Socratic learning assistant
- PhotoQA with generated interactive 2D/3D canvases
- Mistake-book loop, dashboards, and longitudinal insight
- SQLite/PostgreSQL persistence and deployable full-stack application
- Public showcase environment
LalaSystem is an active hyper-personalized, cross-subject, and cross-grade adaptive-learning framework. TaForYou is its current reference application. The public showcase demonstrates Grade 9 mathematics and opens daily from 20:00 China Standard Time (UTC+8), with AI-powered learning features enabled.
Formal archival citation metadata will accompany a future release. For now, cite the repository together with the exact commit or release used:
@software{lalasystem_2026,
author = {Zhou, Zirui and Chen, Kani},
title = {LalaSystem: A Hyper-Personalized Learning Framework},
year = {2026},
url = {https://github.com/HKUST-FintechLab/LalaSystem},
note = {Please include the exact commit or release used}
}LalaSystem is released under the Apache License 2.0. Third-party dependencies, trademarks, and externally sourced assets remain subject to their respective terms.
For any questions, please open an issue in this repository or contact zzhouei@connect.ust.hk.
- Project and collaboration: use the repository issue tracker.
- Organization: HKUST Fintech Lab.
- Security reports: follow
SECURITY.mdand share sensitive details only through private channels.