A personalized AI tutor for programming and mathematics. No catalog, no fixed syllabus — SigmaLoop figures out what you need and generates the course, lessons, coding challenges, math problems, and quizzes on demand, just for you.
Most learning platforms hand you the same content as everyone else. SigmaLoop does the opposite: every course, lesson, and challenge is manufactured by an AI pipeline for one specific learner, and owned by them. There is no shared catalog and no instructor-authored content — the only way material exists is that the system generated it for you.
The atomic experience is a lesson — a short teaching body plus many challenges of mixed kinds. A lesson is complete only when all its challenges pass. A single lesson on recursion might mix a multiple-choice check, a coding challenge, and a math proof.
| 🗣️ Talk to the mentor — a Socratic, autonomous AI tutor. Tell it what you want to learn; it gauges your level and acts: it can read your courses and progress, then create courses, generate lessons, and edit content on your behalf. | 🧭 Guided onboarding — pick topics from a curated library, then answer AI-tailored follow-up questions. The system deduces your goal and generates a curriculum to match. |
 |
 |
| The mentor creates a personalized curriculum on request. | The onboarding wizard — pick what you want to learn. |
The defining engineering opinion of SigmaLoop is how each kind of challenge is graded. The product uses exactly as much AI as it must, and not a drop more:
| Kind | What the AI authors | How it's graded |
|---|---|---|
| PROGRAMMING | prompt, starter code, reference solution, test cases | Judge0 sandbox runs your code against the test cases — deterministic |
| MATH | problem + canonical solution (LaTeX) + rubric | the active LLM returns a structured verdict with a confidence score — AI judgement, confined to where it's genuinely needed |
| MCQ | stem + options (each flagged correct/incorrect) | server-side set-equality — deterministic; option correctness is never sent to the client until you submit |
Programming and MCQ grading stay deterministic (fast, free, never hallucinate a grade). LLM judgement is confined to mathematics — the one place where x^2 + 2x + 1 and (x+1)^2 are the same answer in different clothes — and even there a low-confidence verdict is held for review rather than trusted.
The kind-dispatched workspace: Monaco editor, language picker, Run, and a live output panel.
- Frontend — React 19 + TypeScript + Vite SPA, Tailwind, a clean Linear/Vercel-style design.
- Backend — Node.js + Express + TypeScript REST API (JSend) over MongoDB/Mongoose, JWT auth.
- AI — DeepSeek primary, Google Gemini 2.5 Flash automatic fallback, behind one swappable
AIClientinterface. - Code execution — Judge0 CE sandbox runs AI-generated test cases (Python, C++, Java, JS, TS, Go, Rust).
- Async generation — a curriculum request enqueues a
CurriculumJob; a worker writes theCourse→Lesson→Challengedocuments, so the chat never blocks on generation. - Autonomous mentor — drives the same generation tools through a provider-agnostic
[[ACTION: {…}]]protocol, so a mid-conversation DeepSeek↔Gemini failover never corrupts the tool transcript.
| Repo | What's inside |
|---|---|
| SigmaLoop | Umbrella repo — full technical book (docs/), AWS deployment kit, architecture specs |
| Backend | Node.js + Express + TypeScript API, MongoDB, AI service layer, Judge0 integration |
| Frontend | React 19 + Vite SPA — mentor chat, challenge workspaces, onboarding |