I work at IIT Madras with one of India's largest online degree programs, and I've watched the same story repeat every semester: working professionals enroll with genuine intent, then lose the battle not to the material, but to the calendar. Three weeks of "I'll revise this weekend" quietly becomes exam-week panic.

Meanwhile, every AI study tool I used — NotebookLM, ChatGPT's Study Mode, Gemini's Guided Learning — shared one design assumption: the learner shows up and asks. They are brilliant products you visit. But the science of the forgetting curve says the critical moments happen exactly when you don't visit. The name is the thesis: Padhai (Hindi: studies) + Kairo (Japanese, 回路: circuit) — a closed learning circuit. Say it fast and it's also a friendly Hindi command: padhai karo — "go study!"

I wanted to build the missing half: not a tutor you visit, but an agent that works while you sleep.

🎯 What it does

PadhaiKairo is an autonomous, multi-agent study companion and retention guardian built on one closed loop:

Assess → Model → Plan → Intervene → Re-assess

  1. Ingest anything — paste a YouTube lecture link or your own notes; a multi-tier ingestion pipeline turns them into a grounded study workspace.
  2. Learn Socratically — a tutor agent grounded in your materials guides you with questions and hints instead of handing over answers, with rendered math, code, and tables.
  3. Prove it — a Quizmaster generates 5-question drills with grading rationales; a Flashcard agent builds active-recall decks (Easy/Medium/Hard) with a 3D flip-card UI.
  4. Model memory — every attempt updates a per-concept mastery graph governed by the Ebbinghaus forgetting curve: [ R = e^{-\Delta t / S} ] where (R) is current retention, (\Delta t) is time since last assessment, and (S) is a memory-stability score that grows with each successful review. Mastery you earned but haven't revisited is treated as leaking — not kept.
  5. Plan backward — state your exam date once; the Planner agent backward-schedules interleaved spaced-repetition blocks into your real Google Calendar, densest where decay is worst. Change the exam date and the whole plan re-flows.
  6. Intervene proactively — the Retention Guardian runs autonomous retention scans, flags any concept with (R < 0.40), queues the exact flashcards and drills, drafts a plain-English nudge email — and waits. Nothing sends until you click Approve (human-in-the-loop by design).

Chatbots forget you when you close the tab. PadhaiKairo doesn't let you forget.

Built With

  • docker
  • fastapi
  • gemini-3.5-flash
  • github-actions
  • gmail-api
  • google-calendar-api
  • google-cloud-firestore
  • google-cloud-run
  • google-gemini
  • google-genai
  • google-oauth2
  • python
  • sdk
  • tenacity
  • transcript
  • youtube
Share this project:

Updates

Submission history