Project ideas one prompt can’t build

BeginnerBrowser extension1–2 weeks

YouTube Focus Mode

Hide the distractions on YouTube while you study.

What you’ll build

An extension for Chrome or Edge. Turn on study mode and YouTube hides Shorts, the home feed and the videos suggested next to the one you are watching. A small panel shows your goal for the session and how many minutes you have spent.

Test it with real people. Use it yourself for a week. Then give it to three friends and watch where they get stuck setting it up.

It must do these 4 things

Study mode on and off from the toolbar button.
Hides Shorts, the home feed and the suggested videos.
A goal and a timer for each session that survive closing the tab.
A settings page to choose what gets hidden.

Where one prompt falls short

“Click the next video and the Shorts come back.”

Use AI for any part you like. These are the parts you’ll still have to understand, test and fix yourself.

YouTube changes pages without reloading. The first version usually works on the first video, then the suggestions come back when you click the next one.
YouTube can change its page structure at any time, so hiding things by exact class names can break. You need a sturdier way to find them.
Chrome extensions split code between a background worker and the page. Keeping the timer right when that worker goes to sleep is easy to get wrong.

Build it in 4 steps

Load a basic extension and change one thing on youtube.com.
Hide Shorts and the feed, then keep them hidden as you move between videos.
Add the goal and the timer, saved with the extension's storage.
Add the settings page and test it on a slow laptop.

It is done when

Suggestions stay hidden after 10 video clicks in a row.
The timer is still right after you close and reopen the browser.
A friend installs it from your GitHub page in under 5 minutes.

Go further

A weekly report of the time you saved.
A "5 more minutes" button that locks for an hour after use.
Publish it on the Chrome Web Store.

What you’ll learn

JavaScript
The DOM
Chrome extensions
Browser storage

Download the full brief Submit for a mentor review

BeginnerTelegram bot1–2 weeks

Class Reminder Bot

A bot for your class group that never lets a deadline slip.

What you’ll build

A Telegram bot that sits in your class group. Anyone can add a deadline with a short command, like "/add DBMS file due Friday 5pm". The bot confirms it, lists what is coming up, and posts a reminder the evening before and two hours before.

Test it with real people. Add it to your real class group for two weeks.

It must do these 4 things

Add, list and delete deadlines with simple commands.
A reminder the evening before and two hours before.
Works in more than one group without mixing them up.
Only the person who added a deadline, or a group admin, can delete it.

Where one prompt falls short

“Your server runs on UTC. Your class runs on IST.”

Use AI for any part you like. These are the parts you’ll still have to understand, test and fix yourself.

Servers usually run on UTC and your class runs on IST. Reminders that arrive 5 hours 30 minutes late are the classic first bug.
Free hosting puts apps to sleep. A sleeping bot misses its reminders, so you need a schedule that survives restarts.
People write dates in many ways: "kal", "next Fri", "12/11". Reading them correctly, or asking back when unsure, takes several rounds.

Build it in 4 steps

Create the bot with BotFather and make it reply to one command.
Save deadlines in a database, separately for each group.
Add the reminder schedule and test it across a server restart.
Handle messy dates and confirm with the user before saving.

It is done when

Two weeks in a real group with no missed or repeated reminders.
Deadlines from one group never show up in another.
Restarting the server loses nothing.

Go further

A summary of the week every Sunday evening.
Export deadlines to Google Calendar.
Understand Hinglish date phrases like "parso shaam".

What you’ll learn

Python or Node.js
Bot APIs
Databases
Scheduling
Time zones

Download the full brief Submit for a mentor review

BeginnerWeb app1–2 weeks

Trip Split

Split a trip's expenses and settle up in the fewest UPI payments.

What you’ll build

A web app for a group trip. Everyone adds what they paid for, split equally or by custom amounts. At the end it shows who owes whom with the smallest number of payments, and each payment shows the friend's UPI ID and the amount, ready to pay in any UPI app.

Test it with real people. Use it on a real outing with friends, then compare it with what everyone thinks they owe.

It must do these 4 things

Create a trip and share it with a link.
Add expenses split equally or by custom amounts.
Balances that always add up to exactly zero.
A settle-up list with the UPI ID and amount for each payment.

Where one prompt falls short

“₹100 split three ways must still add up to ₹100.”

Use AI for any part you like. These are the parts you’ll still have to understand, test and fix yourself.

Money stored as decimals goes wrong: ₹100 split three ways must still add up to ₹100, not ₹99.99. You have to store paise and handle the leftover.
"Fewest payments" is a real algorithm question. A method that works for three people can fail for six, so test bigger groups too.
Several friends editing the same trip at once can overwrite each other's expenses.
UPI apps must refuse a payment link opened from another app when the payee is a person (NPCI rule since 1 April 2024). A "Pay" button with the amount filled in won't work.

Build it in 4 steps

Model trips, people and expenses, with amounts stored in paise.
Write the balance and settle-up logic, and test it with 10 made-up trips.
Add the share link and safe editing by several people.
Add UPI IDs and try a ₹1 payment between two phones.

It is done when

Every test trip balances to exactly zero.
A six-person trip made of two groups of three, who only owe each other, settles in four payments.
Two phones adding expenses at the same time don't lose any.

Go further

Expenses in two currencies for a trip abroad.
Scan a bill photo to fill in the amount.
A summary you can share as an image.

What you’ll learn

JavaScript
Data modelling
Algorithms
Testing
UPI payments

Download the full brief Submit for a mentor review

BeginnerPhone app (PWA)1–2 weeks

Attendance Planner

Know how many classes you can miss, under your college's real rules.

What you’ll build

An app you install on your phone's home screen (a PWA). Add your subjects, your timetable and your college's attendance rules. For each subject it tells you how many classes you can still miss, or how many you must attend in a row, and it reminds you the night before a risky day.

Test it with real people. Use your real timetable for a month and check its numbers against your college portal.

It must do these 4 things

Subjects with their own rules: 75%, labs counted separately, medical leave if your college allows it.
Mark a class attended or missed in one tap.
"You can miss N" or "Attend the next N" for every subject.
Works without internet and installs to the home screen.

Where one prompt falls short

“Labs that count double. Holidays. Extra classes.”

Use AI for any part you like. These are the parts you’ll still have to understand, test and fix yourself.

Real rules are messy: labs that count double, holidays, extra classes. The first version gets the simple 75% case right and the rest wrong.
It has to work offline, so your data lives on the phone and must survive the browser clearing it.
Reminders while the app is closed need web push: permission, a service worker and a small server. On iPhone, push works only after Add to Home Screen.

Build it in 4 steps

Write the rules logic and test it with your own past attendance.
Build the timetable and one-tap marking.
Make it installable and make it work offline.
Add the night-before reminder.

It is done when

Its numbers match your college portal for a month.
It opens and works with no internet.
A friend with a different timetable sets it up alone.

Go further

Read the timetable from a photo.
Share a subject's schedule with classmates.
A graph of the whole semester.

What you’ll learn

JavaScript
PWAs
Offline storage
Notifications

Download the full brief Submit for a mentor review

BeginnerBrowser game1–2 weeks

Campus Run

An endless-runner game set in your own college.

What you’ll build

A 2D endless runner in the browser, like the Chrome dinosaur game, set in your college. It works with taps on a phone and keys on a laptop, gets faster over time, and keeps a shared leaderboard for your class.

Test it with real people. Share it in your class group and watch the leaderboard for a week.

It must do these 4 things

Smooth running, jumping and collisions.
Speed that rises the longer you survive.
Works with touch on phones and keys on laptops.
An online leaderboard with names.

Where one prompt falls short

“It runs twice as fast on a 120 Hz phone.”

Use AI for any part you like. These are the parts you’ll still have to understand, test and fix yourself.

Games that move a fixed amount every frame run twice as fast on a 120 Hz phone. Movement has to be based on time, not frames.
Collision boxes that look right but feel unfair. Making it fun takes many rounds of play-testing, not one prompt.
A public leaderboard is easy to fill with fake scores. The server needs some check on every score.

Build it in 4 steps

Draw the player and the ground and get the jump to feel right.
Add obstacles, collisions and rising speed.
Add your college's look and sounds.
Add the leaderboard and basic checks against fake scores.

It is done when

The game runs at the same speed on a 60 Hz laptop and a 120 Hz phone.
10 friends finish runs with no bugs.
A score typed into the browser console by hand is rejected.

Go further

A daily challenge with a fixed obstacle pattern.
Characters you unlock.
A share card with your score.

What you’ll learn

JavaScript
Canvas
Game loops
Server checks

Download the full brief Submit for a mentor review

BeginnerLive web app1–2 weeks

Canteen Board

Today's canteen menu and queue, updated by students live.

What you’ll build

A shared live page for your campus canteen. Anyone can mark what is available today and how long the queue looks. Everyone else sees the change within seconds, and old reports fade out so the page never shows lunch-time news at dinner.

Test it with real people. Ask 15 students to use it at lunch for a week.

It must do these 4 things

Today's menu with "available" or "finished" for each item.
Queue length as short, medium or long.
Changes show up on other phones within 5 seconds.
Reports older than 20 minutes stop counting.

Where one prompt falls short

“Two people report different queues at the same moment.”

Use AI for any part you like. These are the parts you’ll still have to understand, test and fix yourself.

Live updates across phones need a real-time database or sockets, plus hosting that stays up.
Two people report different things at the same moment. You need a rule for which one to show.
Campus Wi-Fi drops. The page must reconnect and catch up without a refresh.

Build it in 4 steps

Build the menu page with a database behind it.
Add live updates and test with three phones at once.
Add the rules for old and clashing reports.
Handle dropped connections, then run the lunch-time test.

It is done when

A change on one phone shows on two others within 5 seconds.
Turning Wi-Fi off and on catches the page up without a refresh.
A week of real use with no wrong "available" items reported by testers.

Go further

Prices and a daily special.
A display mode for a TV at the canteen.
Weekly charts of the busiest times.

What you’ll learn

Real-time databases
WebSockets
Hosting
Testing with users

Download the full brief Submit for a mentor review

IntermediateWeb app + phone scanner3–5 weeks

Fest Check-in

QR tickets for your college fest, with a scanner that works offline.

What you’ll build

A ticket system for a college event. Students register and get a QR ticket. Volunteers at the gates scan tickets with their phone cameras and see a clear yes or no. A ticket can enter only once, even if two gates scan it at the same moment, and the scanner keeps working when the network drops.

Test it with real people. Run it for a club event of 50 to 100 people.

It must do these 4 things

Registration that gives each person a unique QR ticket.
A scanner page that uses the phone camera.
A clear green or red result in under a second.
A live count of people inside, and an export of everyone who came.

Where one prompt falls short

“Two gates scan the same ticket in the same second.”

Use AI for any part you like. These are the parts you’ll still have to understand, test and fix yourself.

Two volunteers scan the same ticket at two gates in the same second. Without the right database check, both let it in.
Gates lose signal. The scanner must work offline and sync later. An offline gate can't see the other gate's scans, so decide how it stops a copied ticket.
A QR code is easy to copy or make up. Tickets need a signature so a made-up code fails.
Camera access behaves differently on Android Chrome and iPhone Safari.

Build it in 5 steps

Registration and signed QR tickets.
The scanner with the camera and an instant result.
The "enter only once" check, tested with two scanners at once.
Offline mode and syncing.
A dry run with friends acting as the crowd.

It is done when

Two phones scanning the same 30 tickets together let each one in only once.
The scanner keeps working in flight mode and syncs afterwards.
A QR code you make by hand is rejected.

Go further

Different ticket types and gates.
Send tickets on WhatsApp or email.
A dashboard for organisers.

What you’ll learn

Full stack
Database transactions
Offline sync
Signed tokens
Camera APIs

Download the full brief Submit for a mentor review

IntermediateAI web app3–5 weeks

Notes to Quiz

Turn your class notes into quizzes that come back before you forget.

What you’ll build

Upload notes or a PDF from your course. The app writes quiz questions from it, shows the line in the notes each answer comes from, and brings questions back on a schedule based on how you answered (spaced repetition), so your weak topics come up more often.

Test it with real people. Use it for one real subject before your mid-semester exams, with two friends.

It must do these 4 things

Upload PDF notes, including scanned ones.
Each question shows its answer and the line in the notes it came from.
A review schedule for each user.
A page that shows your weak topics.

Where one prompt falls short

“Some of the AI's answers will be wrong.”

Use AI for any part you like. These are the parts you’ll still have to understand, test and fix yourself.

Scanned notes are pictures, not text. You need text recognition (OCR) and a plan for bad scans.
AI-written questions are sometimes wrong. Each answer must point to a line in the notes, and users need a way to flag bad ones.
Every question uses API calls, which cost money or use up a free quota. Long PDFs need splitting, saving results and a limit per user.

Build it in 4 steps

Upload and pull out the text, including OCR for scans.
Generate questions with their source lines and check 50 of them by hand.
Build the review schedule.
Measure the cost per user and add limits.

It is done when

At least 45 of 50 hand-checked questions are correct and point to the right line.
A scanned PDF works end to end.
A 30-page PDF costs less than a budget you set and write down.

Go further

Questions in Hinglish.
Quiz battles with friends.
Export to Anki.

What you’ll learn

LLM APIs
OCR
Checking AI output
Spaced repetition
Cost control

Download the full brief Submit for a mentor review

IntermediateReal-time web app3–5 weeks

Live Class Quiz

A live quiz game for 50 phones at once, with a fair timer.

What you’ll build

A host screen on the projector shows each question. Students join from their phones with a room code, answer against the clock and see a live leaderboard. Built for a classroom of 50 or more.

Test it with real people. Run it in a real class or club meeting.

It must do these 4 things

Room codes and joining with a name.
Questions with a countdown.
Scores that reward right and fast answers.
A leaderboard after each question, and rejoining after a dropped connection.

Where one prompt falls short

“50 phones, weak Wi-Fi, one timer.”

Use AI for any part you like. These are the parts you’ll still have to understand, test and fix yourself.

Every phone has a different delay. Timing answers on the phone is unfair and easy to cheat, so the server has to keep time.
50 connections at once, many on weak Wi-Fi. The server must handle drops and rejoins without losing scores.
Two answers arriving in the same millisecond, or the same answer sent twice.

Build it in 5 steps

Rooms and joining over WebSockets.
Questions, timer and scoring on the server.
Reconnecting without losing scores.
A load test with a script that acts as 100 players.
A real class run.

It is done when

A load test with 100 fake players ends with correct scores.
A phone that drops for 10 seconds rejoins with its score.
Changing the phone's clock doesn't change the score.

Go further

Import questions from a spreadsheet.
Team mode.
A replay of the whole game.

What you’ll learn

WebSockets
Server state
Concurrency
Load testing
Deployment

Download the full brief Submit for a mentor review

IntermediateAI web app3–5 weeks

Lost & Found Matcher

Post a photo of what you lost. The app finds likely matches.

What you’ll build

A lost and found for your campus. People post lost or found items with a photo, a place and a time. The app compares photos and descriptions and suggests likely matches, and the two people can chat without sharing phone numbers.

Test it with real people. Collect 40 real photos of everyday items from friends, like bottles, ID cards and earphones, to test the matching.

It must do these 4 things

Post a lost or found item with a photo, place and time.
Ranked list of likely matches.
A private chat between the two people.
Sign-in only with your college email.

Where one prompt falls short

“Matching photos is not one API call.”

Use AI for any part you like. These are the parts you’ll still have to understand, test and fix yourself.

Matching photos needs image embeddings or captions, and then a way to measure whether the top matches are right.
Phone photos are large. Uploading on campus data needs compression and resizing.
Privacy: no phone numbers or room numbers on public pages, and a way to report misuse.

Build it in 4 steps

Posting with compressed photo upload.
Matching, measured on your 40-photo test set.
Private chat.
College-email sign-in and reporting.

It is done when

In your test set, the right item is in the top 3 matches at least 8 times out of 10.
A 5 MB photo uploads as less than 500 KB.
No personal contact details appear anywhere public.

Go further

A notification when a new match appears.
A mode for the campus security desk.
Matching from a description alone.

What you’ll learn

Image embeddings
Vector search
File uploads
Sign-in
Privacy

Download the full brief Submit for a mentor review

IntermediateHardware + web3–5 weeks

Room Monitor

A small sensor kit that tracks your room's heat and noise on your phone.

What you’ll build

An ESP32 board with a temperature and humidity sensor and a sound sensor, placed in your room. It sends readings every minute to your own server. A phone dashboard shows live values and graphs, and alerts you when the room gets too hot or too loud.

Test it with real people. Run it in your room for two weeks.

It must do these 4 things

A reading every minute.
A dashboard with live values and 7-day graphs.
Alerts with limits you set.
Keeps working after Wi-Fi drops and power cuts.

Where one prompt falls short

“Hardware does not follow prompts.”

Use AI for any part you like. These are the parts you’ll still have to understand, test and fix yourself.

Hardware does not follow prompts. A loose wire or the wrong pin needs a multimeter and patience.
Hostel Wi-Fi drops. The board must reconnect on its own and keep the readings it couldn't send.
Sound sensors are noisy. Turning raw numbers into "too loud" needs real testing in your room.
A device that sends data online needs a key, or anyone can send fake readings.

Build it in 5 steps

Read both sensors and print the values.
Send readings to a small API on your server.
Build the dashboard and alerts.
Add reconnecting and saving unsent readings.
Run it for two weeks.

It is done when

Two weeks with less than 1% of minutes missing.
Readings taken during a 10-minute Wi-Fi cut still reach the dashboard.
A request without the key is refused.

Go further

Battery power with deep sleep.
A small screen on the device.
Several rooms on one dashboard.

What you’ll learn

C++ or MicroPython
Sensors
HTTP or MQTT
Time-series data
Device security

Download the full brief Submit for a mentor review

IntermediateDiscord bot3–5 weeks

Coding Club Bot

A daily coding problem for your club, with a leaderboard that checks real solves.

What you’ll build

A Discord bot for your coding club. Every morning it posts a problem at the right level. Members link their Codeforces handle, and the bot checks the Codeforces API to confirm who actually solved it, then updates streaks and a weekly leaderboard.

Test it with real people. Run it in a real Discord server with at least 15 members for three weeks.

It must do these 4 things

A daily problem, chosen by level.
Link and verify a Codeforces handle.
Solves checked through the API, never typed in by hand.
Streaks, a weekly leaderboard and admin commands.

Where one prompt falls short

“How do you prove a Codeforces handle is yours?”

Use AI for any part you like. These are the parts you’ll still have to understand, test and fix yourself.

Proving a handle belongs to the person: you need a check, like asking them to submit a compile error to a problem you pick.
The Codeforces API allows one call every two seconds. Checking 50 members takes 100 seconds, so you need a queue and saved results.
The bot must keep running, and must not post twice after a restart.

Build it in 5 steps

A bot that posts a problem on a schedule.
Handle linking with verification.
Solve checks within the API limits.
Streaks, the leaderboard and admin commands.
Three weeks in a real server.

It is done when

A post every day for three weeks, never twice.
Someone else's handle cannot be linked.
API calls stay under the limit you measured.

Go further

Contest reminders.
Team battles between years.
Hints that open after a few hours.

What you’ll learn

Bot APIs
External APIs
Rate limits
Scheduling
Databases

Download the full brief Submit for a mentor review

AdvancedInfrastructure6–10 weeks

Your Own Deploy Platform

Push a GitHub repo, get a live link. Your own small Vercel.

What you’ll build

Connect a GitHub repo. Your platform builds it inside a container, gives it a subdomain like app.yourname.dev, shows the build logs live, and lets you roll back to an earlier version.

Test it with real people. Deploy five different real projects: a static site, a React app, a Node API, a Python API and one that fails to build.

It must do these 4 things

A push to GitHub starts a build.
Builds run in separate containers with time and memory limits.
Each app gets its own subdomain with HTTPS.
Live logs, and rollback to the last working version.

Where one prompt falls short

“You are running other people's code on your server.”

Use AI for any part you like. These are the parts you’ll still have to understand, test and fix yourself.

You are running other people's code. Without separation and limits, one bad build can take down the whole server.
Sending each subdomain to the right container, with HTTPS certificates, needs a reverse proxy you set up and understand.
The old version must keep serving until the new one is healthy, or every deploy causes downtime.

Build it in 5 steps

Build one repo by hand in Docker and serve it.
Trigger builds from GitHub pushes.
Subdomains, HTTPS and the reverse proxy.
Live logs, limits and rollback.
Switch versions with no downtime.

It is done when

All five sample projects deploy, and the broken one fails with a clear log.
A build stuck in an endless loop is stopped by the time limit.
A check sending one request a second sees no failures during a version switch.

Go further

Environment variables and secrets.
A preview link for every pull request.
A usage dashboard.

What you’ll learn

Linux
Docker
Networking
Reverse proxies
Security

Download the full brief Submit for a mentor review

AdvancedAI voice app6–10 weeks

Hinglish Voice Helpdesk

Ask your college questions out loud in Hinglish, and get answers with the source.

What you’ll build

A voice assistant for your college. Students ask in Hinglish, like "exam form ki last date kya hai?". It turns speech into text, finds the answer in your college's real documents (notices, rule book, timetable), answers out loud and shows the document it used. If the documents don't cover it, it says so.

Test it with real people. Collect 60 real questions from students, with the right answers from the documents.

It must do these 4 things

Speech in and speech out, in Hinglish.
Answers only from the documents, with the source shown.
"I don't know" when the documents don't cover the question.
Half of all replies start within 3 seconds.

Where one prompt falls short

“It will make up dates unless you stop it.”

Use AI for any part you like. These are the parts you’ll still have to understand, test and fix yourself.

Speech-to-text can get Hinglish wrong. You have to test models and fix the common mistakes.
Finding the right paragraph in 200 pages of PDFs is hard, and the wrong paragraph gives a wrong answer.
The model will make up dates unless you make it answer from the sources and check that it did.
Every part adds delay: speech-to-text, search, the model and text-to-speech.

Build it in 4 steps

Collect the documents and the 60-question test set.
Search over the documents, measured on the test set.
Answers with sources and "I don't know".
Speech in and out, then cut the delay.

It is done when

At least 50 of the 60 test questions answered correctly, with the right source.
Questions outside the documents get "I don't know".
Half of all replies start in under 3 seconds.

Go further

Voice notes on WhatsApp.
Hindi in Devanagari script.
An admin page to add new notices.

What you’ll learn

Speech models
Retrieval (RAG)
Embeddings
Evaluation
Latency

Download the full brief Submit for a mentor review

AdvancedReal-time app6–10 weeks

Code Together

Code together in the browser like Google Docs, then run it safely.

What you’ll build

A code editor in the browser where several people edit the same file at once and see each other's cursors. A Run button runs the code on your server in a sandbox and streams the output to everyone in the room.

Test it with real people. Pair-program with friends on real DSA problems for a week.

It must do these 4 things

Several people editing at once with no lost text.
Live cursors with names.
Run Python or JavaScript with time and memory limits.
Rooms with share links.

Where one prompt falls short

“Two people type in the same line at once.”

Use AI for any part you like. These are the parts you’ll still have to understand, test and fix yourself.

Two people typing in the same line at once. Simple syncing loses text; you need a method built for this, like CRDTs.
Running strangers' code: an endless loop or a "delete everything" command must not touch your server.
Someone goes offline, keeps typing and comes back. Their changes must merge cleanly.

Build it in 5 steps

An editor in the browser with rooms.
Live editing and cursors with a CRDT library.
The sandbox for running code.
Offline edits and merging.
A week of real pair-programming.

It is done when

Three scripts typing at random for 5 minutes end with identical copies.
An endless loop is stopped by the time limit.
Code in the sandbox cannot read files outside it.

Go further

Voice chat in the room.
More languages.
A replay of the whole session.

What you’ll learn

CRDTs
WebSockets
Sandboxing
Containers
Editor libraries

Download the full brief Submit for a mentor review

AdvancedFull stack + payments6–10 weeks

Campus Marketplace

Buy and sell used books and cycles on campus, with payment held until pickup.

What you’ll build

A marketplace for your campus. Sellers list items with photos. Buyers pay through Razorpay in test mode, so no real money moves. The money is held until the buyer confirms pickup, and either side can raise a dispute.

Test it with real people. Run it with test payments among 20 friends, with real listings.

It must do these 4 things

Listings with photos and search.
Checkout in test mode.
Payment held until pickup is confirmed.
Refunds, disputes and sign-in with a college email.

Where one prompt falls short

“The same payment message arrives three times.”

Use AI for any part you like. These are the parts you’ll still have to understand, test and fix yourself.

Payment providers send confirmation messages (webhooks) late, twice or out of order. Without care, one payment is counted twice.
Two buyers pay for the last copy of a book at the same moment.
If the buyer never confirms pickup, who gets the money? You need clear rules and time limits.

Build it in 5 steps

Listings, photos and search.
Test-mode checkout and webhooks.
Holding money, confirming pickup and refunds.
A ledger that records every rupee.
Disputes and a test run with friends.

It is done when

The same webhook sent three times records one payment.
Two buyers paying for one item at once ends with one sale and one automatic refund.
The ledger balances to the last rupee.

Go further

Ratings for buyers and sellers.
Chat between them.
Delivery inside campus.

What you’ll learn

Payments
Webhooks
Idempotency
Transactions
Ledgers

Download the full brief Submit for a mentor review

AdvancedSearch engine6–10 weeks

College Search

Search every page and PDF on your college website, ranked properly.

What you’ll build

A crawler that reads your college website and its PDFs, an index of everything it finds, and a search page that puts the right notice or rule first in under a second, even with typos.

Test it with real people. Write down 50 real questions students search for, and the page that should come first for each.

It must do these 4 things

A polite crawler that follows robots.txt, limits its speed and finds new pages.
Text from PDFs, including tables.
Ranked results with a short snippet.
Typo tolerance, and a refresh every day.

Where one prompt falls short

“Getting the right notice first, not just any page with the words.”

Use AI for any part you like. These are the parts you’ll still have to understand, test and fix yourself.

College websites are messy: broken links, the same page at three addresses, scanned PDFs.
Ranking is the hard part: the right notice first, not just any page that contains the words.
You can't tell if search is good without your own test set and a score.

Build it in 5 steps

Crawl the site politely and store pages and PDFs.
Pull out clean text and remove duplicates.
Build the index and the first ranking.
Score it on your 50 questions and improve the ranking.
Daily refresh and a fast search page.

It is done when

For your 50 questions, the right page is first at least 35 times and in the top 3 at least 45 times.
New notices show up within 24 hours.
Search answers in under 300 milliseconds.

Go further

Questions in Hinglish.
Search by meaning, not just words.
Alerts when a new notice matches a saved search.

What you’ll learn

Crawling
Parsing
Indexes
Ranking
Evaluation

Download the full brief Submit for a mentor review

AdvancedAI agent6–10 weeks

AI Bug Hunter

An agent that uses a website like a person and reports what's broken.

What you’ll build

Give it a website and a goal, like "sign up and add an item to the cart". It drives a real browser with Playwright, tries to finish the goal, and writes a bug report with the steps and screenshots whenever something fails.

Test it with real people. Build a small shop app and hide 10 bugs in it yourself. Then let the agent loose.

It must do these 4 things

Runs a goal from start to end in a real browser.
Bug reports with steps and screenshots.
Never reports the same bug twice.
A limit on steps and cost for each run.

Where one prompt falls short

“A slow page is not a bug.”

Use AI for any part you like. These are the parts you’ll still have to understand, test and fix yourself.

The agent gets stuck or goes in circles on pop-ups and slow pages.
False alarms: a slow page or an expected error is not a bug. You need checks before it reports.
You need a score: hidden bugs found against false reports.

Build it in 5 steps

Build the test shop and hide 10 bugs.
Let the agent drive the browser for one goal.
Add bug reports with screenshots and remove repeats.
Add the step and cost limits.
Score it and cut the false alarms.

It is done when

It finds at least 7 of your 10 hidden bugs.
Fewer than 3 false reports per run.
Every run stays inside the step and cost limits.

Go further

Run it on every pull request.
Compare screenshots between versions.
Suggest a likely fix with each report.

What you’ll learn

AI agents
Tool use
Playwright
Evaluation
Cost control

Download the full brief Submit for a mentor review

Project ideas one prompt can’t build.

Pick your level.

Beginner

Hide the distractions on YouTube while you study

A bot for your class group that never lets a deadline slip

Split a trip's expenses and settle up in the fewest UPI payments

Know how many classes you can miss, under your college's real rules

An endless-runner game set in your own college

Today's canteen menu and queue, updated by students live

Intermediate

QR tickets for your college fest, with a scanner that works offline

Turn your class notes into quizzes that come back before you forget

A live quiz game for 50 phones at once, with a fair timer

Post a photo of what you lost. The app finds likely matches

A small sensor kit that tracks your room's heat and noise on your phone

A daily coding problem for your club, with a leaderboard that checks real solves

Advanced

Push a GitHub repo, get a live link. Your own small Vercel

Ask your college questions out loud in Hinglish, and get answers with the source

Code together in the browser like Google Docs, then run it safely

Buy and sell used books and cycles on campus, with payment held until pickup

Search every page and PDF on your college website, ranked properly

An agent that uses a website like a person and reports what's broken

Get it reviewed by a mentor.