Utkarsh Classes· EdTech
Teacher-level doubt solving, at 48 crore queries a year
We built a doubt engine for 4 lakh paid students that answers photo and text questions at senior-teacher quality, and cut the cost per request by around 70 percent.
- Applied AI
- Retrieval-Augmented Generation
- Multimodal Input
- Cost Optimisation
70%
Lower cost per request
The challenge
Utkarsh has roughly 4 lakh paid students with access to a premium app, and they wanted instant doubt solving inside it. Students needed to type a question or upload a photo of a maths or science problem and get an answer explained the way a senior Utkarsh teacher would explain it. At three to four questions per student per day, that works out to around 4 crore queries a month, or 48 crore a year. Answering that at quality was pushing teacher workload and cost per request up.
What we built
We built a doubt-solving engine designed for quality at national scale.
- Ingestion of a ready database of around 2 lakh questions into a retrieval system
- A daily pipeline that pulls current information, including general knowledge and current affairs from newspapers, so answers stay up to date
- Multimodal input: a student can snap a photo of a problem or type it out
- Explanations pitched at senior-teacher level, tuned for the competitive maths and science courses the platform runs
- A stable system engineered to hold at 48 crore queries a year
The results
The engine answers at teacher quality, takes routine doubt load off teachers, and increased demand for the premium app. The headline is cost: operating cost per request fell from around 70 paise to roughly 20 to 22 paise, a cut of about 70 percent, across 48 crore queries a year.
At a glance
- /Cost per request cut from ~70 paise to ~21 paise, around a 70% reduction
- /48 crore queries a year handled at stable quality
- /A 2 lakh question bank plus a daily current-affairs pipeline
- /Photo and text doubts across maths and science
- /Teacher workload reduced and premium demand increased
Built with
- Retrieval-augmented generation
- Multimodal input
- Data pipelines
- LLM cost optimisation
- Scalable inference
More stories
Other engagements
Work with us
We want your hardest problems.
We collaborate with ambitious organisations ready to move beyond AI experimentation. If you're looking to transform how your business operates, we'd love to build with you.