Unclaimed

Finds the scholarships an Indian student can still claim, checks every rule with the exact sentence it came from, and catches when websites get the deadline wrong. Nemotron + Tavily.

Comment

Inspiration

On 28 October we looked up one scholarship, Karnataka's post-matric scholarship for Backward Classes, and found three different last dates. The department's own page said 31 August. Buddy4Study, where most students look, said 30 September. The state portal's notice board, in Kannada, said 31 October. Two of the three said it was closed; it closes on Saturday. (The department page was updated to 31 October a day later; Buddy4Study still says 30 September.) Multiply that by every department, the National Scholarship Portal, AICTE and the foundations, rules that change every year and income limits that differ by caste group (₹2.5 lakh for Category-1, ₹1 lakh for 2A/3A/3B), and students find out about the scheme they qualified for after it closed. Meanwhile the money goes unclaimed: government figures given in Parliament show post-matric scholarship beneficiaries among SC, OBC, EBC and DNT students fell from 45.45 lakh to about 24.5 lakh over four years (The Wire, 12 Mar 2026).

What it does

Eight facts in, or one tap on a demo student. Under a minute later: every scholarship found, each eligibility rule ticked ✓, crossed ✗ or marked ? against you with the sentence it came from, which ones close soonest, where a website disagrees with the official page, one document checklist, what to do this week, and a calendar file with reminders.

How we built it

A nine-step agent, each step labelled in the interface with the model, tool or code doing it:

  • Nemotron 3.5 Lightning plans the searches.
  • Tavily Search runs them on the open web, restricted to gov.in/nic.in, and on the state scholarship portal. Tavily Extract reads the full pages.
  • Nemotron 3 Super (JSON mode) extracts each rule with a word-for-word quote, including from Kannada pages. The portal's notice board has a fixed format, so code reads its dates.
  • Code merges copies of the same scheme across pages, takes the newest official date and shows the older one; code decides eligibility; a verifier rejects any rule whose quote isn't on the page and flags when a secondary site disagrees with the official one.
  • For rules still unsettled, the agent searches government sites again for that one rule and re-checks.
  • Nemotron 3 Ultra ranks what's left and writes the plan.

Next.js on Vercel, Server-Sent Events for the live timeline, a spend guard, recorded real runs for the demo students, 23 unit tests and a labelled accuracy test.

Challenges

  • Three dates for one scheme. Deadlines get extended and department pages lag. We group copies of a scheme by level and department (names never match) and trust the newest dated official notice.
  • The model mixed up years on the portal's notice board, which lists every year's dates in the same Kannada sentence. A parser reads it exactly.
  • Nemotron thinks by default. Lightning spent its whole token budget thinking and returned nothing; Super took up to a minute per page. Switching thinking off for extraction took a run from 182 s to under a minute.
  • "Official" isn't enough. Our first ranking put a Manipur page first for a Karnataka student, and a look-alike domain (sspscholarshipkarnataka.in) mimics the real portal.
  • Money decisions shouldn't come from a model, so the model only reads; code decides; every rule shows its source.

Accomplishments

  • On hand-labelled pages, over three runs: 34 of 36 rules and 14 of 15 deadlines read correctly, 0 invented rules.
  • It catches the real Buddy4Study date error for the BCW scheme and the Minorities Welfare one (30 Sept vs the official 30 Nov).
  • A full run costs about $0.015 (₹1.5) and takes 30–75 s.

What we learned

Token Factory's OpenAI-compatible API made switching models trivial, and JSON mode was reliable once thinking was off. Nemotron 3 Super reads Kannada notices and copies quotes in the original script. Turning reasoning off where the job is copying, and on where it's judging, was the biggest speed and cost win. Tavily's include_domains is what makes official portals findable at all. And the hardest part of this problem isn't AI: it's that the official sources disagree with each other.

What's next

More states' notice boards, WhatsApp reminders, and a college scholarship-cell view that runs the check for a whole class.

Limitations

They're real and published on purpose: see the README. The demo students are fictional; their one-tap results are recorded real runs from 29 Oct, labelled as such, and "Run it live" runs the whole agent again.

Built With

  • nebius-token-factory
  • nemotron-3-super
  • nemotron-3-ultra
  • nemotron-3.5-lightning
  • next.js
  • nvidia-nemotron
  • tavily
  • typescript
  • vercel
Share this project:

Updates

Submission history