Skip to content

Lumina EdTech · Education

RAG tutoring assistant for 200k learners

A curriculum-grounded AI assistant answering with 94% measured accuracy.

Industry
Education
Services
RAG Development, AI Development, LLM Applications
Stack
Python · LangChain · Claude · PostgreSQL
Timeline
4 months
Region
India / SEA

94%

answer accuracy on golden set

1.4M

learner questions answered/mo

58%

support tickets deflected

$0.011

cost per answered question

01

The challenge

Lumina's learners asked hundreds of thousands of questions monthly that outpaced human tutor capacity. Generic chatbot pilots hallucinated beyond the curriculum and damaged trust with parents and schools.

02

The solution

We built a retrieval-augmented assistant grounded strictly in Lumina's curriculum: syllabus-aware retrieval, grade-level answer calibration, and hard refusal outside curriculum scope. Every answer cites the lesson it draws from, and low-confidence retrievals route to human tutors.

03

Architecture & engineering

An ingestion pipeline keeps 40,000 curriculum documents synced and chunked; hybrid retrieval with a reranker feeds Claude through a versioned prompt system. An evaluation harness scores 1,200 golden questions on every release, gating deploys on accuracy regressions.

04

The outcome

The assistant now answers 1.4M questions a month at 94% measured accuracy, tutor escalations dropped by more than half, and per-question cost is a rounding error against tutor hours.

Sydova treated accuracy like an engineering metric, not a marketing claim. That rigor is why schools trust the product.
Priya RamanCPO, Lumina EdTech

Services used

How we did it

Start a similar project

Facing a challenge like Lumina EdTech's? We'll show you how the same discipline applies to your roadmap.