FLN assessments are the standard diagnostic for grade-level literacy and numeracy gaps in Indian schools — they're the correct test, just operationally expensive. A teacher administers, listens, and scores by hand for every student, which means the period becomes mostly scoring overhead, not teaching.

Students complete the assessment with real-time speech analysis scoring their response; the result posts back to a Java / Spring Boot service that aggregates across the cohort. A React frontend gives teachers a single per-classroom view instead of 40 paper sheets. The backend was sized for 100+ concurrent assessments so a whole school can test together.

1. Real-time speech analysis, not post-hoc review

   Doing the speech analysis in real time meant teachers get scoring instantly instead of queueing audio for later. 85% scoring accuracy against ground truth was the bar that made it a viable substitute for manual scoring rather than just an aide.

   decision · 01
2. Java / Spring Boot for concurrency at school scale

   100+ concurrent assessments at peak (a full school cohort testing at once) is the kind of load Spring Boot handles cleanly. A JVM stack gave reliable concurrency without bespoke tuning and slotted into deployment muscle memory the team already had.

   decision · 02
3. React + PostgreSQL for per-classroom rollups

   Teachers needed a per-classroom rollup, not per-student score reports. A React frontend over PostgreSQL-backed aggregations gave each teacher one view of where the cohort actually was, which was the operational win the product was optimising for.

   decision · 03

Pushing speech analysis to be lighter and on-device would have made deployment to low-connectivity schools easier — the network was the real constraint, not compute. And the dashboard could have surfaced what students struggled with, not just how many, with another iteration of aggregation logic.