Lumi
Lumi is the AI platform for education. Teachers create assignments and upload materials once; students work through them with an AI tutor that only answers based on the teacher's actual content, generate quizzes and study flashcards, and get personalized support based on their learning profile. Coordinators and principals see how each student is doing in real time with an actionable traffic-light indicator, manage digital report cards and keep AI costs under control. Built for K-12 schools, special-needs and inclusive education programs, technical schools and private secondaries.
What is Lumi
Lumi is an AI platform for education that replaces the passive LMS and manual grading with conversational assistants that tutor the student while they work, automatically grade with explanations, and generate actionable reports about each child's academic standing — without making things up: assignment and material chats are anchored to the content the teacher uploaded.
Everything is configured from each school's admin panel — from the AI provider (OpenAI, Anthropic, Groq, Ollama or any compatible endpoint) to the tone and rules of every one of the 13 system prompts, budget caps, the academic calendar and the grading thresholds. No code required.
Why Lumi
1:1 tutoring at any hour
The student moves forward whenever they can — the AI chat is available 24/7, focused on what the teacher uploaded.
AI that does not make things up
When an assignment or material has attachments, the AI answers exclusively based on that content. Anchored to the teacher's material.
Real teacher productivity
The per-subject pedagogical assistant creates assignments, grades submissions, analyses students and generates adaptive activities — operated through chat.
Inclusive education that actually works
Three accommodation levels, a structured learning profile per student and a simplified UI with adaptive activities generated by AI.
Actionable visibility
A green / yellow / red / grey traffic light per student, recomputed monthly, with three concrete bullets to act on.
Frictionless report cards
Teachers fill in, admin reviews, two-phase publishing, automatic PDFs. The manual gradebook spreadsheet is over.
Predictable costs
Every AI call is logged with tokens and cost. Dashboards break it down by school, teacher, student and source.
Isolated multi-tenant
Each school has its own logical instance, AI provider, academic calendar, prompts and admins. No data crossover.
Who it is for
Lumi is built for educational institutions that want to bring AI in a structured way — with traceability, cost control and respect for the teacher's content:
If your institution still manages students, subjects and report cards in spreadsheets and loose documents, Lumi unifies the flow in a platform with AI anchored to the teacher's official content.
Typical use cases
Student solves an assignment with the AI tutor
Receives the assignment with attachments, opens the chat, asks questions. The AI only answers from the teacher's content, no hallucinations. When done, the student submits and the entire conversation is logged.
Student studies a material via chat, quiz or flashcards
Over a PDF, YouTube video or audio uploaded by the teacher, the student chats, generates a self-graded quiz with explanations, or a flashcard deck for active recall.
Teacher creates and assigns work via chat
Asks the subject's pedagogical assistant: "build me an activity on X for group 7A due Friday". The assistant creates it, assigns it and notifies the students.
Teacher analyses students and grades with AI
Requests a performance analysis per student, receives actionable points and a PDF for the file. Grades submissions with AI-suggested feedback and manual adjustments.
Coordinator detects students in red
Opens the dashboard, sees the school or group traffic-light distribution, filters the reds, reads the three bullets per student and assigns actions to teachers or tutors.
Inclusion: adaptive activity for one student
The teacher of a student with significant accommodations asks the system for a "shape-matching" activity on a topic. The AI generates it with assets (images, audios) that get reused across schools.
Platform capabilities
Per-role and per-subject assistants
Lumi exposes three conversational assistants in the app, each with its own scope and permissions. The philosophy is focused assistants instead of one generic assistant that has to know everything.
- Lumi (floating chat): available to student, teacher and admin from any screen. Queries system data based on role — between 14 and 22 read-only tools.
- Per-subject pedagogical expert: for the teacher, with 23 tools that include reads and writes (create assignments, grade submissions, analyse students).
- Curriculum chat: lets the teacher explore the uploaded subject syllabus through natural-language questions.
- SSE streaming: responses appear word by word, with status events while the assistant queries data.
In-task AI tutoring grounded in the material
When an assignment or material has attachments, the student's chat runs in source-only mode: the AI is instructed not to make things up and to answer exclusively from the content the teacher uploaded. What the student learns is what the teacher posted, nothing more, nothing less.
- PDF, DOCX, TXT and audio as attachments. The system extracts the text and passes it as context to the model.
- Post-submission justification: after sending the answer, a second chat verifies that the student actually understood what they submitted.
- Student audio: the student records a spoken answer, it is transcribed automatically with Whisper and tied to the submission.
- Full traceability: every conversation between student and tutor is recorded and available to the teacher.
Active learning with materials
A material is not a PDF to read and forget. On every material the student can chat, generate quizzes and create flashcard decks — all generated by AI from the teacher's content.
- Auto-generated MCQ quizzes: five questions per quiz, instant grading, explanation for every right and wrong answer.
- Study flashcards: concept / definition cards for active recall.
- Study podcasts: the teacher can generate an audio of a material via TTS (ElevenLabs) and students listen to it inside the app.
- Cross-material history: the student reviews every quiz and deck in a single view.
Inclusive education adaptive activities
Lumi supports three accommodation levels per student
(NONE / NON_SIGNIFICANT / SIGNIFICANT) and a
dedicated flow for the significant ones: a simplified UI
(/student/lite) without sidebars, with one activity at
a time and immediate feedback.
- Nine adaptive activity types: visual selection, matching, sequencing, classification, fill-in, puzzle, silhouette, guided reading, guided writing.
- AI generation with the student profile: the teacher defines topic and goal, the AI builds items considering the learning profile and the available assets.
- Reusable asset repository: images, audios and words shared across schools. The AI reuses them or generates them on demand and persists them as global so the next school that needs them gets them for free.
- Enriched prompts for non-significant cases: the structured student profile is injected into every AI prompt, improving the accompaniment without changing the base content.
Student learning profile
Each student with an accommodation has a structured profile built from a guided AI interview between the school admin and the system. The AI asks questions, the admin answers what they know about the student, and on closing the conversation the normalized profile is extracted.
- Fields covered: reading/writing, preferred comprehension mode, response style, step tolerance, strengths, interests, barriers, supports, summary.
- Injectable into any prompt: the placeholder
{STUDENT_PROFILE}is replaced automatically in any prompt that declares it. - Editable by hand: the admin can adjust the fields after the interview, no need to start over.
Student traffic light monthly recompute
A green / yellow / red / grey indicator per student and per subject, computed by AI considering absolute performance, personal trend and proactivity. Each computation comes with three actionable bullets and is stored to track evolution over time.
- Monthly cron: on the 1st at 3 AM the system processes every student in the school.
- Evolution heatmap: a temporal view with the colour chain of each student over the year.
- Aggregated distributions: by subject, by group, by teacher, by school. Filters to list students in each colour.
- Bi-weekly notification to the teacher with a summary of changes in their subjects.
Digital report cards with two-phase publishing
Report cards are managed with a coordinated flow: the admin creates the period, teachers fill in behaviour, performance and written remark in their subjects, and publishing happens in two steps to avoid a half-finished card reaching the student.
- Per-teacher / per-subject close and reopen: each teacher closes when they are done. The admin can adjust afterwards with no restriction.
- Two-phase publishing: first visible to teachers (review), then finalized and visible to students.
- Automatic PDF with the per-subject breakdown, the period's assignment grades and the written remark.
- Automatic notifications to teachers when the period is published and to students when it is finalized.
Notifications and messaging
Five automatic notification types for students (new assignment, new material, grading, report card published, others) and two for teachers (new submission, report card published). Direct messaging 1-on-1 and groups inside the same school.
- Unread count in each role's sidebar.
- Direct messaging between any pair of users from the same school.
- Groups with ADMIN/MEMBER roles and polling for near-real-time updates.
Per-school multi-LLM no lock-in
Each school configures its own AI provider. Lumi talks to any endpoint compatible with OpenAI's chat completions API — which today covers most of the ecosystem. If you want to switch providers tomorrow, it is a change in the admin UI.
- Supported providers: OpenAI, SAIA, Anthropic via LiteLLM, Groq, Ollama, or any other compatible endpoint.
- Basic and advanced models: per school, with independent input/output prices, to optimize cost per flow.
- Whisper for audio transcription (spoken submissions) — managed or external.
- ElevenLabs for study-audio generation (TTS).
- Image generation on demand for adaptive-activity assets.
- 13 system prompts configurable at three levels: global default, per school, per subject.
Identity and multi-school national ID + JWT
Each user is identified by their national ID (one user, one ID) but can hold different roles in different schools — the same person can be a teacher in one and a parent or admin in another. Login resolves which school to enter based on role and permissions.
- Native multi-role: the
user_schoolstable maps (user, school, role). No duplicate accounts. - Subdomain per school: each institution can have its own URL (
school.lumi.imasdev.com). - Force password change on first login or when an admin resets credentials.
- Login rate limiting: the security filter protects against brute force without penalizing the legitimate user.
Observability and traceability
Every student-AI chat, every tool the pedagogical expert invokes, every LLM call and every HTTP request is recorded with tokens, cost and latency. The "the bot told my kid something weird" complaint is resolved by opening the exact conversation.
- Full conversations: assignments, materials, assistants — everything persists in the database.
- Tokens and cost per call: the central
llm_usage_logwith breakdowns by school, teacher, student and source. - API observability: latency, errors, throughput, slow endpoints. Per school or global.
- AI analysis history: every analysis (submission, student, teacher performance, engagement) is stored and downloadable as PDF.
AI cost control no surprises
The AI bill is the first thing that scares people away from LLMs. In Lumi every call is priced in USD using the configured provider's rates, and the admin gets dashboards with every breakdown to see where the budget is going.
- Total and daily cost in USD, with a bar chart for the selected range.
- Breakdown by source: assignment chat, pedagogical assistant, analysis, quiz generation, traffic light, and more.
- Breakdown by teacher and by student: useful to spot atypical usage or training needs.
- Basic and advanced models: each flow picks the right model to optimize cost without losing quality.
Integrable into any institution
Lumi runs on a standard stack: Java 21 + Spring Boot 3, PostgreSQL with Flyway migrations, React 19 + Vite on the frontend, all packaged with Docker. It runs from a modest VPS to an infrastructure with a managed database and several nodes behind a load balancer.
The LLM client is generic — any endpoint compatible with OpenAI's chat completions API works, which includes OpenAI, Anthropic via LiteLLM, Groq, Ollama, SAIA and internal proxies. Authentication is JWT with national ID; login rate limiting and multi-tenant isolation come out of the box.
Dedicated deployment per client: the instance belongs to the school or to the school network. No data crossover, no artificial limits, no arbitrary "plans". The database, the AI provider and the storage belong to whoever deploys.
Differentiators
- AI grounded in the teacher's material. When there are attachments, the tutor answers only from that content. No hallucinations, no improvising, no talking about things the class did not study.
- Real inclusive education. Three accommodation levels, a structured per-student profile, a simplified UI, and nine types of adaptive activities generated by AI with assets reused across schools.
- An actionable traffic light, not another dashboard. Green/yellow/red/grey per student with three concrete bullets to act on, recomputed monthly with an evolution heatmap.
- Multi-LLM with no lock-in. Every school picks its provider, its basic and advanced models and its prices. Switching provider is a change in the UI.
- Configuration is 100% UI. The 13 system prompts, the prices, the academic calendar, the grading thresholds and the AI providers are all editable from the admin panel.
- Full traceability. Every chat, every tool invoked, every LLM call and every HTTP request is persisted with tokens, cost and latency. Essential to answer questions from families or principals.
- Dedicated deployment. Your instance, your database, your AI provider. The student-tutor conversation does not feed any shared model.