Case Study 01 — AI EdTech · 0→1
Listen2RE
An AI-augmented audio learning platform that turns wasted commute hours into UPSC progress.
Problem
UPSC/MPSC preparation takes years, the syllabus is enormous, and most aspirants hold full-time jobs. The standard solutions — books, coaching, YouTube — all demand screen-on, focused attention that working aspirants simply cannot give.
Working aspirants commute 45–90 minutes daily in conditions where reading is impossible — that time produces zero progress.
Existing audio content was low-quality YouTube: slow pacing, no structure, nothing designed for audio-native learning.
Aspirants knew exactly what to study. The constraint was never knowledge of the syllabus — it was usable time.
Research
Field interviews with working aspirants aged 24–34, employed full-time, with 1–3 hours/day for prep. Mapped their actual day hour-by-hour instead of asking what features they wanted.
“I know what I need to study. I just can't find the time to sit and study it.”
- 01Commute time was the single largest block of recoverable learning inventory — and it was 100% unused.
- 02Aspirants had tried podcasts and abandoned them: content read aloud from documents doesn't work for ears.
- 03Decision fatigue was real — large content libraries caused skipped sessions, not more engagement.
Insights
Insight 1 — Passive commute hours are wasted learning inventory — the product is reclaiming time users already spend.
Insight 2 — Audio-native ≠ text-to-speech. Content designed for ears has different sentence structure, pacing, and signposting than content designed for eyes.
Insight 3 — For habit-forming behavior, consistency beats choice. One great daily session outperforms an infinite library.
Strategy
Position Listen2RE not as a podcast or TTS app, but as an AI content system that ingests dense UPSC material and produces structured, audio-native daily sessions matched to each learner's commute.
Product Requirements
- Daily 15–25 min audio session, pushed at 6:30 AM, matched to the learner's current topic
- LLM pipeline: topic extraction → concept summarization → audio-script formatting → key-term callouts
- Quality gate: LLM-as-a-Judge pre-filter plus 10-minute human spot-check before publish
- Engagement loop: session rating, streaks, and replay with topic override
- Mobile-first web app with PWA caching for offline listening
Prioritization — RICE, weighted by habit impact
Every candidate feature scored on Reach × Impact × Confidence ÷ Effort, with a habit-impact multiplier. Features that strengthened the daily loop (push timing, streaks, voice quality) consistently out-scored content breadth. The library expansion everyone asked for scored lowest — and was cut.
PRD — North-star metric
Completed listening sessions per learner per week — not downloads, not signups.
PRD — Explicit non-goal
Listen2RE is not an on-demand podcast library. One daily session, sequenced for the learner.
PRD — Quality guardrail
No AI-generated session ships below an 85% LLM-as-a-Judge pass threshold. Human review on every flag.
Execution
Built the full content system end-to-end: ingestion, LLM processing on the Claude API, dual-layer quality review, voice synthesis, and CDN delivery — orchestrated with n8n, instrumented with Mixpanel.
System Architecture
User Journey
- 16:30 AM — daily session lands, matched to current topic
- 2Commute — 15–25 min structured audio, paced for listening
- 3Key-term recap — callouts reinforce retention at session end
- 4One-tap rating — feeds the quality loop
- 5Streak + progress — visible momentum across the syllabus
Key Decisions & Trade-offs
AI generates, human spot-checks — not the reverse
Trade-offOccasional quality misses a human-first flow would catch. Accepted because the user feedback loop surfaces issues fast — and production effort dropped 60%.
Daily push over on-demand library
Trade-offLess perceived user control. Offset with topic overrides and replay. Habit strength won: 71% completion.
Mobile web first, native app later
Trade-offNo offline listening at launch — the #1 feature request. Shipped 8 weeks instead of 6 months; PWA caching closed the gap in a later iteration.
Metrics
Lessons
Personalization earlier. Topic-sequencing by individual progress should have been month 2, not month 8.
Community sooner. Learners wanted to discuss sessions — a simple thread per episode would have lifted retention for 14 months.
Evals are pre-launch infrastructure. Shipping without systematic LLM evaluation cost us early retention. Never again.
Voice quality was the real product. The voice-pipeline upgrade beat every feature shipped that quarter.