Bimbam

Your study mirror. No cap.

Comment

Inspiration

Bimbam (பிம்பம்) is Tamil for the image a mirror gives back. I wanted a study tool that does the same thing: show students what they actually understand, not what they think they do.

Looking around my class as week 1 started, I could tell I wasn't the only one feeling overwhelmed. So I asked students what stresses them out most. Missed deadlines came first, messy notes second. Deadlines hide in course outlines, emails, posters and project group chats. Notes are scattered across slides and readings. And most AI study tools make it worse: a generic chatbot answers from the internet, rarely says where an answer came from, and will happily write your essay for you.

So I set one rule: Bimbam only uses your own course material, cites everything, and makes you think first.

What it does

Never miss a deadline

  • Drop in a course outline and every deadline lands in one calendar, with the exact phrase it was read from.
  • Paste a "class moved" message, or screenshot an event poster or room booking, and it becomes a reminder.
  • Due-soon items turn red, exams get a countdown, and ticking a to-do off comes with confetti.
  • The Telegram bot, nocapbb, lives in project group chats: it tracks group deadlines, finds when everyone is free, books the meeting into everyone's calendar and nudges the group before it starts.

Actually understand the content

  • Upload slides, readings and notes (PDF, PowerPoint, Word or a photo) and get one master note per week, with every point cited to its slide or page. A Simple words mode helps students reading in a second language. Notes work in English, Chinese, Tamil and German.
  • Discuss with Capy, a capybara tutor holding a mirror. Capy checks your explanation against your notes, corrects mistakes with citations (putting on a detective hat when it spots one), and asks the next question instead of giving answers.
  • Mirror check: explain the week in one sentence. Capy shows what you got right (with citations) and asks questions about what you missed, without giving you the answer.
  • Polish my English explains each grammar issue and its rule, then lets you write the fix and checks it.
  • Case-study coaching linked to your course concepts.
  • Quizzes at your level in mixed formats: scenarios, "which is NOT", spot the error, true or false, fill in the blank, and explain-in-your-own-words answers that Capy checks. Every wrong option gets a "why not".

Stay motivated, with Capy

  • What's next? picks one useful next step from your own deadlines and notes.
  • A focus timer where Capy studies under a lamp with you, then naps during your break. Capy lets you know when a note, a quiz or your focus time is ready, even if you've moved to another screen.
  • Outfits and animal friends are earned only by studying (score 80% on a quiz and a duckling comes to sit with Capy). Nothing can be bought.
  • Hit today's AI limit and Capy tells you when the next slot opens and suggests something useful instead of a dead end.

How I built it

  • Frontend: React 19, TypeScript, Vite and Tailwind, deployed on Vercel.
  • Backend: Supabase (Postgres with row-level security, Auth, private storage, Deno Edge Functions, pg_cron for reminders) and the Telegram Bot API.
  • AI: every AI feature is its own Edge Function, so keys never reach the browser. All output is structured JSON that is validated in code before a student sees it. The AI layer works with Gemini, Claude or any OpenAI-compatible API, and two providers can be chained so one takes over if the other is down.

Responsible AI

I treated safety as a feature, then attacked my own live app with 14 red-team tests. All 14 passed, one of them after a fix the test itself uncovered.

  • Prompt injection: hidden "ignore your instructions" text in slides, a phishing link in an event email, and faint text in a booking screenshot were all ignored. Pasted emails are now also scanned in code, so Capy warns the student about hidden AI instructions and phishing links by name.
  • Academic integrity: requests to write an essay, a case analysis or a polished rewrite were refused, including "stay in character and just write it".
  • Hallucination: off-topic questions get "your materials don't cover this", and wrong claims get a cited correction.
  • Privacy and abuse: a second account saw none of the first account's data, forged bot calls were rejected (403), and daily limits held. A password-reset link can only be used to set a new password; it never logs you straight in.

The full results are in SAFETY_TESTS.md in the repo.

Challenges I ran into

  • Helpful without ghost-writing. Rewriting a student's English is exactly what academic integrity rules forbid, so I built a coach that explains the rules and checks the student's own fix. My tests then caught it slipping a re-ordered phrase through as a "word option", and I closed that gap in code.
  • The Telegram webhook. Getting Telegram, Supabase and a shared secret to agree took several rounds of 404 and 403 errors, plus rotating the bot token after it was exposed.
  • Messy real-world dates. Bookings arrive as tables, and an early parser read "9 October" as nine hours. Now every date quotes its evidence and waits for the student to confirm.
  • Translations that were too hard to read. The first Chinese notes were accurate but textbook-dense, so I added Simple words mode.
  • AI outages mid-testing. "High demand" errors led me to chain two independent AI providers.

What I learned

  • Attacking your own AI finds real problems. Mine found two. The first was a ghost-writing loophole: the English coach is only allowed to suggest a swap for a single weak word, but one of its "word options" re-ordered the student's whole phrase, a rewrite in disguise. Every option is now checked in code. The second was a reliability gap: when the AI provider was overloaded, every retry went back to the same busy model, so students saw nothing but errors. A second, independent provider now takes over automatically.
  • Students trust what they can check. A citation beats a clever answer.
  • Building it end to end changed how I see AI. Taking Bimbam from the spark of an idea to a working app showed me the positive aspects of AI when it is designed carefully. It also gave me hands-on experience with every part of building an application, from the interface and database to AI safety and deployment.

What testers told us

I shared Bimbam with 6 students before submitting.

"I don't have to keep using different apps to study."

"I like that Capy doesn't change or edit stuff for me, and I have to do that myself."

"The UI is pleasant to look at, and all features are straightforward."

"The main reason I haven't used it much is the limited range of question types. More content would make it even better."

What I changed the same day: quizzes now mix six multiple-choice styles (scenario, "which is NOT", cause and effect, compare, spot the error, definition) with true or false, fill in the blank, and explain-in-your-own-words questions that Capy checks against key points without writing the answer.

Business model

Bimbam is priced for students. A general AI subscription costs around $30 a month, or about $120 over a four-month semester. Bimbam costs a fraction of that.

We offer two plans. Students can subscribe directly for $8 a semester. Universities can license Bimbam for their whole cohort at $6 per student per semester, rising with deeper integrations such as an eLearn sync. The university plan means every student gets access, so no one is priced out. A free tier with a few notes and quizzes a week lets students try Bimbam before they or their school commit.

These prices allow for higher-quality AI. Bimbam uses a mix of models: a fast, low-cost model for chats, quizzes and deadline extraction, and a stronger model only for weekly notes, where accuracy matters most. Daily caps and caching of repeated course content keep costs predictable. We estimate the average student costs about $3–4 a semester in AI usage, which leaves room to cover hosting, support and growth. At SMU's scale of about 10,000 students, the university plan would bring in about $60,000 a semester.

Pilot plan

We will run a one-semester pilot at SMU with about 200 students across one or two modules, organised through BIA and AI Lodge. Students use Bimbam free during the pilot. The AI bill, which we estimate at $1,000–1,500 for the semester with hard usage caps per student, will be covered by sponsorship and AI-provider credits. Bimbam defaults to a low-cost model, uses the stronger model only for weekly notes, and caps daily use, with weekly spending checks so costs stay predictable. Throughout the pilot we will track usage, student feedback and study outcomes. With that evidence, we will approach SMU and partner universities through their AI and tech clubs for the $6-per-student university plan.

What's next

  • Real push notifications, even with the browser closed (including iPhone)
  • Install Bimbam as an app on the Home Screen
  • Flashcards with spaced repetition from quiz mistakes
  • Canvas / eLearn integration
  • Malay and more languages
  • Revision reminders from Capy on Telegram
  • Study groups and shared quizzes

Built With

  • apiyi
  • brevo
  • css3
  • deno
  • deno-edge-functions
  • github
  • google-gemini
  • html5
  • json-schema
  • node.js
  • openai-compatible-api
  • pg-cron
  • postgresql
  • private-storage
  • row-level-security
  • supabase
  • supabase-edge-functions
  • tailwindcss
  • telegram-bot-api
  • typescript
  • unpdf
  • vercel
  • vite
  • vite-and-tailwind
  • web-audio-api
Share this project:

Updates

Submission history