Second Reader

First-pass marking for handwritten answer scripts. It reads first; the teacher decides.

Live demo: https://second-reader-two.vercel.app · Built for the Nebius x NVIDIA Global AI Hackathon · MIT licence

A teacher photographs each page of a student's script on their phone. NVIDIA Nemotron 3.5 Super VL on Nebius Token Factory reads the handwriting question by question, Nemotron 3 Ultra proposes marks strictly against the teacher's rubric with a one-line reason, and plain code caps marks, rounds to half marks, adds the total and highlights anything blank, unclear or uncertain. Nothing is final until the teacher approves it; then the marks register downloads as CSV.

---

The problem

Indian college faculty mark stacks of handwritten internal-test scripts by hand, often over a weekend. Errors happen at scale: Bangalore University blacklisted 50 evaluators for erroneous marking of answer scripts (Deccan Herald, 9 Jan 2025), and totalling mistakes are common enough that universities offer re-totalling separately from re-evaluation. Marking needs a teacher's judgement, but reading forty scripts' worth of handwriting, checking each answer against the scheme of valuation and totalling it is slow, repetitive work.

What it does

First-pass marking

1. Set up the test: questions, maximum marks and the rubric, typed or filled from a photo of the answer key.
2. Snap each page of a script (resized on the phone to about 1600 px before upload). Roll number only, no names.
3. Read (Nemotron 3.5 Super VL): each answer transcribed, with how sure it is.
4. Mark (Nemotron 3 Ultra): a proposed mark and a one-line reason per answer, against the rubric.
5. Audit (code): caps at the maximum, half-mark rounding, totals, and flags: left blank, hard to read, marker unsure, over max, not read, extra answer.
6. Review: flagged answers first, confident ones folded away, the page photos above with tap-to-zoom, −½/+½ buttons or typing, then Approve and Next script. Snap the next script while one is still being read.
7. Marks register (CSV) once every graded script is approved; mark one question across the whole class; the question that lost the class the most marks; time and cost per script.
8. Check against my own hand marks: type your paper marks and see how far the first pass was, and whether the big gaps were already flagged.

Totalling check Photograph a script you've already marked. Super VL reads only the numbers you wrote; code adds them up and compares with your written total, and flags marks over the maximum or not in half marks.

Which model does what, and why

Step	Model / tool	Why
Read handwriting, answer keys, margin marks	nvidia/Nemotron-3.5-Super-VL, thinking off	Vision; transcription is copying, not reasoning (~5 s per page)
Propose marks	nvidia/Nemotron-3-Ultra-550b-a55b, thinking on	The judgement step: partial credit against a rubric (~4 s)
Caps, rounding, totals, flags, agreement, totalling check	TypeScript (auditor.ts, classwork.ts)	Arithmetic and money-like decisions belong in code

All calls go through Token Factory's OpenAI-compatible API; images are sent as base64 data URLs.

Measured so far

• 29 Oct, real calls: Super VL transcribed a test page exactly in 5.4 s (including "1/R = 1/4 + 1/6 = 5/12"). The full read → mark → audit took 10 s. Ultra gave Q1 2/2, docked Q2 a mark for the missing heating-effect definition (and accepted Imax/1.414 as Imax/√2), gave Q4 2/2, and the audit flagged the blank Q3.
• Cost: about $0.006 per script (≈ ₹0.5) at list prices.
• Synthetic benchmark (29 Oct, real calls): five test scripts rendered in different handwriting-style fonts on ruled paper with tilt and noise (scripts/synth-scripts.ps1), each answer written to earn a known mark against the sample rubric (neat and complete; messy and partial; one question skipped and one wrong; a forgotten 1/R inversion; crossed-out first attempts). Graded through the real route (scripts/bench.mts): within ½ mark on 20 of 20 answers, exact on 19, all five totals within 1 mark; the skipped answer was flagged as blank and the crossed-out attempts were ignored. These are not real student handwriting; that test is still to come.
• Totalling check, real call: on a script marked by hand with red margin marks 2, 1½, 3, 2 and a written total of 9, it read every mark and reported "Written total is 9, but the marks add up to 8.5".
• Answer key from a photo, real call: a typed scheme of valuation became the test's four questions, marks and marking points exactly.
• Latency varies with load: a page read took 5 s in the morning and 45–130 s in the afternoon of 29 Oct; a full script 10–45 s.
• Agreement with a teacher's hand marking on real scripts: TODO.

Run it locally

Key in ../.env (outside this repo): NEBIUS_API_KEY=...

powershell -ExecutionPolicy Bypass -File .\start.ps1           # http://localhost:3001
powershell -ExecutionPolicy Bypass -File .\start.ps1 -Fixture  # layout-only sample reading, no API calls, development only
npm test                                                      # auditor, agreement, class insight, totalling check, mark parsing, cost
npm run smoke:vl -- path\to\page.jpg                          # one Super VL call on a photo

Limitations

• Not tested on real student handwriting yet at the time of writing: the test page used a handwriting-style font. Messy handwriting, diagrams and crossed-out text will lower legibility; low-legibility answers are flagged for the teacher to read.
• The marker can be strict. On a live test it gave 1/4 where a teacher might give 2, because the rubric asked for fuller descriptions. That's why every mark is a proposal and the agreement check exists.
• Totalling check and answer-key reading have each been run for real once, on rendered test pages.
• Photos stay in the browser and are sent only to be read; nothing is stored. Use dummy roll numbers.
• The spend guard is in-memory per server instance: a brake, not a wall.

Licence

MIT — see LICENSE.