Source: https://velofy.co/research/feedback-loop/

Published: 2026-08-23T00:00:00.000Z
Updated: 2026-08-23T00:00:00.000Z

[RESEARCH](https://velofy.co/research/) / ARCHIVE

# The feedback loop exam prep never had

A field dossier on UPSC exam prep: coverage is free on YouTube; the product is rubric-scored Mains evaluation in minutes, adaptive revision, open calibration.

Velofy · 23 Aug 2026

From the archive. This page preserves the original publication; product plans and capabilities may have changed. See [current open source work](https://velofy.co/open-source/).

This is a field dossier: our internal working notes on the government-exam preparation market. Decisions marked D1–D5 are derived direction for heyIAS, not shipped fact.

**The outcome an aspirant buys is coverage × feedback quality × retention.** What aspirants pay for today is structure and evaluation.

## Four truths about the market

1.  **H1: Outcome = coverage × feedback quality × retention.** The scarce inputs are feedback and measurement; neither requires more video.
2.  **H2: The scarce resources were expensive because they needed human graders.** Personalized feedback, Mains answer evaluation, and honest progress measurement were priced like scarce labour.
3.  **H3: Prep spans one to three years.** Attrition is driven by unmeasured progress and isolation, not by a lack of videos.
4.  **H4: Aspirants distrust marketing but trust analytics computed from their own data.**

The market shape: UPSC CSE draws 0.5–1M serious applicants a year; state PCS exams add several million across UPPSC, BPSC, MPPSC, RPSC and the rest; SSC, banking and railways widen the funnel tenfold. Spend is lumpy: offline coaching at ₹1–2L a year for a minority, test series at ₹3–15k, while the mass market self-studies on YouTube, PDFs and Telegram.

## The triangle nobody owns

| Incumbent | Model | Where it breaks |
| --- | --- | --- |
| **Vision IAS / Drishti / Vajiram** | Content broadcast + human-evaluated test series | Evaluation takes days to weeks, costs heavily, feedback stays generic |
| **Testbook / Adda247** | MCQ practice engines at scale | Excellent objective-question plumbing, essentially zero Mains capability |
| **Physics Wallah UPSC** | Lecture model, aggressively priced | Same H1 flaw: sells coverage in a market where coverage is free |

What no incumbent owns is a triangle: fast rubric-based descriptive evaluation, adaptive revision, and published calibration.

## Five decisions the truths force

1.  **D1: Lead with Mains answer evaluation.** Photo or PDF upload, evaluated against a published rubric (structure, content, examples, word discipline) within minutes, percentile against cohort, model-answer delta view.
2.  **D2: Build the Navgati layer.** After every mock, update three things: a topic-wise mastery map, a spaced-repetition queue built from personal errors, and a rank trajectory band with a confidence interval.
3.  **D3: Calibration as brand.** Each cycle, publish “our projected band contained the rank X% of the time.” No competitor dares publish theirs.
4.  **D4: Free artifacts carry distribution.** PYQ deep-analysis reports, one free diagnostic full mock with a weak-topic map, and a Telegram bot that answers PYQ doubts with citations to standard sources.
5.  **D5: Retention comes from data gravity, not streaks.** The lock-in is two years of personal error data no rival can import.

## The eight-week MVP

1.  Ingest ten years of UPSC Mains GS papers, official rubrics and topper copies.
2.  Evaluation engine: LLM scoring against rubric dimensions, strict JSON output, confidence thresholds; low-confidence answers route to a human queue.
3.  Cohort percentile service, seeded by launching with free mocks until N is real.
4.  Revision scheduler: an SM-2 variant over error tags.
5.  Mobile-first web app with Hindi and English copy, plus a Telegram bot surface.

## Pricing: a direction, not a price list

Our working notes sketch three tiers: free diagnostic mock as the hook, a low monthly subscription for unlimited evaluation plus the revision queue, and a full-season program whose projected-rank guarantee ties refund terms to completing the plan. The benchmark is price against test-series spend, not coaching fees.

## Risks

*   **R1: LLM evaluation variance.** Counter with rubric decomposition, temperature zero, dual-pass scoring, sampled human audits, and a published agreement rate.
*   **R2: Copycats and content piracy.** Content is not the moat. Per-user data plus calibration history is; neither can be scraped or copied.
*   **R3: Exam-cycle seasonality.** Smooth it with the PCS calendar, these exams run all year, and later verticals (SSC, JE) sharing the same evaluation engine.

### The short version

Sell what coaching cannot: minutes-not-weeks rubric-scored Mains evaluation, a revision queue grown from your own errors, and a rank band whose accuracy we publish every cycle.

This dossier applies the exam-preparation case derived in [three markets, one thesis](https://velofy.co/research/first-principles/).

[← FastAPI, done properly](https://velofy.co/research/fastapi/)[Three markets, one thesis →](https://velofy.co/research/first-principles/)