AvsarLens

INSPIRATION In small-town India, opportunity still arrives on paper. Research shows 80% of rural students report no awareness of pre-matric scholarship schemes, and thousands of scholarships go unclaimed every year simply because students do not know they exist. The institution — the school notice board — is the single largest source of scholarship information, yet the paper notice assumes the reader can decode government language, remember deadlines, and assemble documents alone. First-generation learners and their parents lose at exactly that step. Existing AI scholarship tools are English-first search portals built for college students; none of them reads the physical notice where the information actually lives.

I am 15, from Bokaro, Jharkhand. I am that student. I built the tool I wished my school's notice board came with.

WHAT IT DOES

1. Capture any notice — camera or upload, any language, bad print, glare, angles.
2. Verified extraction (Gemini 2.5 Flash, BYOK): fixed output card — WHAT IS THIS / WHO CAN APPLY / DEADLINE (+ days left) / DOCUMENTS NEEDED / HOW TO APPLY / EXTRA DETAILS.
3. "Can I apply?" — three questions generated from the notice itself, giving LIKELY ELIGIBLE / MAYBE — ASK SCHOOL / NOT ELIGIBLE, always ending with "final confirmation: contact the office listed on the notice."
4. Tracker — saved notices with countdown and status (New / Asked school / Applied / Missed), stored only in your browser.
5. Print view — a clean one-pager to hand to a parent or teacher.

It refuses to guess. No deadline printed means "VERIFY MANUALLY". Unreadable fields get amber low-confidence flags. Two notices in one photo get flagged, never merged. A missed opportunity hurts; a fabricated deadline hurts worse.

HOW IS THIS DIFFERENT FROM JUST ASKING AN AI CHATBOT?

• Output: a raw chatbot gives a different essay every time; AvsarLens gives the same verified card, every time.
• Deadlines: a chatbot can guess ("usually ends in March"); AvsarLens reads only what is printed, and flags when absent.
• Eligibility: a chatbot gives confident prose; AvsarLens uses rules from the notice plus a 3-question check with an ask-school fallback.
• Follow-through: a chat closes and everything is lost; AvsarLens keeps a deadline tracker with countdowns.
• Privacy: your photo on their servers vs. your key, your browser, nothing stored remotely.
• Built for: people who write perfect prompts vs. a 14-year-old with a cracked phone camera.

AI is the engine. AvsarLens is the vehicle.

HOW WE BUILT IT Solo, as an AI-native builder: I direct AI tools, review every line, and own the architecture, safety rules and testing. Vanilla HTML/CSS/JS single-page app — no framework, no backend, no account. Gemini 2.5 Flash (fixed model, Flash-only fallbacks) via the user's own API key stored in localStorage; the browser calls the API directly and no photo ever touches a server I control. Inline SVG icons, zero external resources, full-English UI. Deployed on Render.

Honest timeline: building started Oct 15, 2026 — before the competition window opened. The repository's commit history shows the entire journey with real dates; nothing was back-dated or faked.

CHALLENGES WE RAN INTO

• Designing for refusal: the easy product says something confident about every photo. The honest one must detect what it cannot read and flag it — deadline-never-invented tests are part of the eval harness, not an afterthought.
• Our own eval caught a real bug: the first full run scored 13/14. The failure was fixture-11 — two separate ads in one photo. The output was honest (both deadlines verbatim, nothing merged) but came back high-confidence with a double deadline and no flag: a student could follow the wrong one. We added a prompt rule (multiple notices means never merge, flag and lower confidence), shipped it the same day, and re-ran the full suite: 14/14. The harness fails us first — before any judge can.
• Real-world capture conditions: notice boards behind glass, folded circulars, faded newsprint, mixed Hindi/English. The 14-fixture synthetic set exists to test exactly this — torn, faded, glare, blur, handwritten, mixed-language.
• Question generation: "Can I apply?" questions must come from the notice itself, never from a generic template, or the check becomes theatre.

ACCOMPLISHMENTS THAT WE'RE PROUD OF

• A verification-first extraction pipeline with fixed output pattern and honest confidence flags
• An in-browser honesty eval harness (14 synthetic fixtures, runs with your own key, ships in the repo): run 1 = 13/14 (93%), caught a real edge case, fixed same day — full-suite re-run = 14/14 (100%)
• Deadline-never-invented behaviour verified on the no-deadline fixture: it refused to guess
• From zero to deployed live product in three days, solo, at 15 — including repo repair and deployment done from a phone via Termux git
• A privacy architecture a parent can understand in one sentence: "your key, your browser, nothing leaves your phone except the API call you own"

WHAT WE LEARNED

• The best AI products are not the ones that know everything — they are the ones that clearly say what they do not know.
• Research before building: one search killed my first idea (board-photo-to-notes has 5+ competitors) and the data pointed here — an empty format space on top of a documented 80% awareness gap.
• Testing the refusal path matters more than testing the happy path.
• An eval harness is not a deliverable you add at the end — it is the thing that makes every other claim in this description trustworthy.

WHAT'S NEXT FOR AVSARLENS

• Regional-language voice summary for parents who cannot read the card (opt-in, still no data leaves the device beyond the user's own API call)
• School mode: a teacher photographs the board's notices once, students' trackers sync via QR
• Offline fallback: on-device OCR for basic fields when there is no network
• Publishing the fixture set + eval harness so other student builders can test their own extraction tools

AI-USE DISCLOSURE (STEMergent requirement) AI tools were used for code generation, research assistance, fixture generation and demo narration under my direction. I reviewed and tested every component; the architecture, safety rules ("never invent a deadline"), eval harness and product decisions are mine. All synthetic demo notices are marked SAMPLE with invented details. Significant AI use is disclosed here per competition rules.

Built With

• css
• html