AI-powered UPSC Mains prep with intelligent answer evaluation and personalized learning.
UPSC Mains answer writing is notoriously hard to self-assess: answers are handwritten, evaluated against a subjective rubric, and good feedback usually means paying for a mentor's time. Aspirants also need fast access to previous year questions with model answers, and topic-specific enrichment — facts, Supreme Court judgements, committee reports — that takes hours to compile manually. Pulse IAS wanted a platform that could give students AI-speed feedback without losing the credibility of expert review, backed by a content bank curated by their own faculty.
We scoped and built this as an MVP: a working product covering answer evaluation, PYQ practice and AI-assisted content tools, validated end-to-end before a wider rollout. The core idea was a rubric-driven evaluation pipeline — students upload a handwritten answer sheet, get it digitized, and choose between an instant AI score or a slower hybrid path where a human mentor reviews the AI's output. Two further AI features, a model-answer generator and an auto-tagged enrichment bank, were grounded in Pulse IAS's own curated study material rather than open-ended generation. The engagement covered product design, the interactive prototype, and the initial backend/AI build; Pulse IAS paused the project before it reached public release.
Client
Web app for submitting answers, practicing PYQs, generating model answers and browsing enrichment content.
API
Core product, evaluation and content services, plus real-time status push for results.
Data
Structured storage for users, submissions, evaluation records, PYQs and enrichment content — all tagged by GS paper, subject and theme.
AI evaluation service
OCR digitization of handwritten answer sheets, followed by LLM-based scoring against a structured rubric.
Async infrastructure
Background processing for OCR, AI scoring and mentor-review queuing, so submissions never block the API.
Mains answers are handwritten, so every submission has to be digitized before it can be scored — and scoring itself runs through a fast AI-only path or a slower AI+mentor path, depending on what the student picks.
The student uploads a scanned or photographed answer sheet, which goes straight to a private S3 bucket via a presigned URL.
A queued OCR job extracts the handwritten text and auto-detects the source question, GS paper and theme, so nothing needs manual tagging.
The digitized answer is scored by the LLM evaluation service against a five-part rubric — Directive, Introduction, Structure, Holistic content, Answer/conclusion — producing per-dimension scores and specific feedback.
For hybrid evaluations, the AI's pre-score and feedback are queued for a human mentor, who layers personalized commentary on top before results are released.
Once scoring — and mentor review, if selected — completes, the result is pushed to the student in real time, instead of polling for a report that can take up to 48 hours.
Problem 1
UPSC Mains answers are handwritten under exam conditions, so there's no clean digital text to evaluate — a photo or scan is all a student has.
Solution
We put OCR digitization ahead of evaluation: uploaded sheets are converted to text and the source question, GS paper and theme are auto-detected before scoring starts.
Impact: Students get evaluated on the same handwritten sheets they'd produce in the real exam, with no manual setup required.
Problem 2
A pure AI score isn't trusted by serious aspirants, but pure human mentor review doesn't scale or stay affordable.
Solution
We designed a hybrid evaluation path: AI scores every submission instantly against the rubric, and students can additionally route it to a mentor queue for expert commentary layered on top of the AI's structured breakdown.
Impact: Students choose speed (instant AI) or depth (AI + mentor in 24–48 hrs) based on what a specific answer needs.
Problem 3
Generic model answers don't help aspirants — they need answers built from Pulse IAS's own curated facts, judgements and committee reports, not just whatever an LLM already knows.
Solution
The model-answer generator and enrichment bank are grounded in a tagged content store: topics are auto-classified to GS paper, subject and microtheme, and the LLM composes introductions, analysis and conclusions from retrieved Pulse IAS source material rather than open-ended generation.
Impact: Generated content stays anchored to the syllabus and to Pulse IAS's own material, not generic AI output.
Problem 4
AI evaluation and generation calls carry real inference cost, so unmetered usage isn't sustainable.
Solution
A credit ledger meters every AI-driven action — a full AI evaluation, a hybrid AI+mentor evaluation, a model-answer generation, or a single-dimension quick check each cost a different amount — tied to subscription plans.
Impact: Usage-based costs stay predictable for the business while students get pay-for-what-you-need flexibility.
OCR-to-evaluation, the hybrid AI+mentor workflow, the model-answer generator and the enrichment bank were all built and functioning end-to-end.
Handwriting OCR accuracy and rubric-based LLM scoring were proven feasible before further investment, rather than assumed.
The rubric-driven, hybrid AI+human scoring pattern is applicable beyond UPSC Mains, to any subjective, long-form answer evaluation.
The engagement concluded at MVP stage; Pulse IAS did not proceed to a public launch.
Building something similar?
Let's Talk