Inspiration
Every student lives in one chaotic class group, where homework, deadlines and announcements get buried under hundreds of messages, greetings and memes. A global study of 10,500 office workers across 11 countries found people lose more than 3 hours a day to repetitive digital tasks — in India, 80% said it blocks their real work. Students repeat their own version of that loop every single day. I watched homework get missed in my own class group and realised: chat apps don't organise study tasks, note apps need manual typing, and no AI tool reads a class group and tells you what to do. So I built one.
What it does
HomeworkLens reads class-group screenshots (or pasted chat text) with Gemini Vision and turns the chaos into a clean checklist. It extracts homework, announced exams, things to bring and important notices — while filtering out greetings, memes and small talk, and even reports how many messages it ignored. It converts relative words like "tomorrow" or "Friday" into real dates with urgency badges, sorts tasks by priority, and shows a focus strip with everything due today or tomorrow plus a countdown to the next exam. A second tab grades answers: paste a question, your answer and the maximum marks, and the AI scores it like a strict examiner, with a verdict, strengths, missing points and an improvement tip. Output works in English, Hindi or Hinglish; the checklist can be copied, shared to WhatsApp or printed; the last 15 scans are saved as history; and a live demo mode works with zero setup. There is no backend and no signup — the API key and all data stay in the browser.
How we built it
I am a solo, AI-native builder: I designed the product, wrote the extraction prompts and the UI spec, directed AI coding tools to generate the code, and verified every feature end-to-end before shipping. The stack is vanilla HTML, CSS and JavaScript with zero dependencies, calling the Gemini multimodal API with structured JSON output, and localStorage for settings, checklist state and history. The submission video was also produced with AI tooling — slides rendered programmatically, narration synthesised, and real-world evidence (the global workforce study and India's UDISE+ student data) woven into the story. It is deployed live on Render.
Challenges we ran into
Getting a vision model to return reliable structured JSON from noisy, multi-screenshot chats required strict schema prompting plus safe parsing with fallbacks. Converting relative dates like "kal" or "Friday" into real ISO dates took careful prompt engineering. Fitting a complete narrated demo — proof, product, demo, code and impact — under the strict 3-minute video limit forced ruthless editing. And keeping everything client-side for privacy, while handling API keys responsibly, shaped several design decisions. On the human side: I balanced this build with school and two other hackathons in the same week.
Accomplishments that we're proud of
A working, deployed product with a clean professional UI and zero emojis; a feature set that goes beyond scanning (deadline intelligence, focus strip, AI grading, bilingual output, history, live demo mode); a fully evidence-backed pitch backed by real data; and doing all of it solo at 15 as an AI-native builder — the idea, the product, the proof and the pitch.
What we learned
AI makes you fast, but verification is everything: strict schemas, error states and fallbacks are what make an AI product trustworthy. Scope discipline wins — one core loop, polished perfectly, beats ten half-built features. And the idea is only 5%; the other 95% is knowing exactly what to build, and proving it works.
What's next for HomeworkLens
WhatsApp-forward intake (forward a chat export instead of screenshots), reminder notifications, a teacher mode for announcements, more languages, and offline PWA support. I plan to pilot it with real class groups in my school and open-source the learning
Log in or sign up for Devpost to join the conversation.