Product builder · AI for learning · Yale SOM MBA '28
I build voice-first AI products that help people practice hard skills.
Five years building learning products with India's public school systems, after five years in coal mining operations and strategy at Tata Steel. Now an MBA student at Yale, building AI practice simulations.
- 49 institutes in the spoken-English pilot, against 18 planned
- 44% of pilot learners retried a passage without being asked
- ~50,000 classroom observations logged in a state's new system
Featured · Oct 2026 · Built for an Anthropic hackathon
Abhyas Studio
First-person job simulations on a smartphone, where India's school mentors rehearse their whole job: teaching a class, giving a teacher feedback after a lesson, and leading their schools. AI models voice the characters; rules I wrote in code decide how they react.
- Mentoring · 2 visits live
- Teaching · 1 of 9 live
- Leadership · 4 of 9 live
- 7 of 8
- mentors named a change for their next real visit
- 21 of 21
- spoken turns of the feedback conversation worked in the field
- Rs 14
- (about 16 US cents) of AI and speech per practice visit, against a hard cap of Rs 25
1Teach a class of 41 children by voice
2See the moment from a child's side
3A teacher who opens up only when you earn it
4Coached on your own words, then retry
Screens from the latest build, Immersive mode. Characters are fictional and illustrated.
Products
Shipped with teams in India's public education systems
Products I owned or co-owned. Each page states my exact role, the decisions I made, and which numbers are results and which are only signals.
- Deployed
Wadhwani AI · 2025-2026
SEAP · Spoken English practice
Spoken English practice for vocational students. I defined success as coming back, not signing up, and the pilot data refocused 2.0 on habit and on making sentences.
ROLE Owned strategy, pilot metrics and 2.0
49institutes in the pilot I designed, against 18 planned
Read the case study - Live since 2022
Central Square Foundation · 2022-2023
State classroom-observation system
An offline observation tool built in three weeks for the mentors' training, a free live dashboard, then a role-based state app for teachers, mentors and officials.
ROLE Product lead · built the offline tool and dashboard myself
~50,000classroom observations logged by mentors
Read the case study - Piloted
Central Square Foundation · 2025
Sakhee · AI teacher coach
Teachers record a lesson's audio on a phone and get kind, specific feedback on the teaching steps they did and missed. I chose the problem and led the product.
ROLE Product lead · set up the AI vertical
53% → 74%of key teaching steps used, first to last recording (50 teachers)
Read the case study - At scale
Wadhwani AI · 2025-2026
AI reading assessment · ORF
A child reads aloud and speech AI scores every word. The program has assessed more than 9 million children in two states; I worked on taking it to new ones.
ROLE Senior Program Officer · state expansion
5 morestates signed agreements while I worked on expansion, Sep 2025 to Jul 2026
Read the case study
Built for myself
Small tools and games I built end to end, with AI coding agents.
Gated
the next case stays locked until I do the practice the grader set
In personal use
Case Room
A voice case-interview simulator for practicing structured problem solving out loud. An AI interviewer runs the case, a second model grades it, and practice gates the next case.
Private, friends only
Chitralahari
The Telugu movie quiz nights I run with friends, as a phone game: song rounds by ear, live rooms, and questions friends write in a spreadsheet.
How I work
Four working rules, each shown on a product page
- 01Teachers' reactions, not the dashboard, set what the AI optimizes.In Sakhee's first test, 15 of 43 of the model's judgments wrongly told teachers they had skipped a step they did, and teachers named it as their worry. So the model was tuned to avoid that error first.See it
- 02Code decides; AI models only talk.In Abhyas, the teacher's trust, every score and what she may reveal are computed by rules in code. A plain chatbot gave her secret away in 3 of 3 runs; in 13 live visits, Abhyas never did.See it
- 03A simple offline version comes before the real app.A prototype that worked without a signal let mentors record classroom observations from their first week of training, months before the state's app launched.See it
- 04Every number says what it counts and how many.Each page marks its claims by status, such as live, measured or piloted, gives small counts as "k of n" with a source, and names the next test.See it
How I got here
From 650 meters down to the classroom
2016
Coal face
Led 130 people underground and commissioned a continuous miner in one of India's deepest coal mines
2019
Control room
Strategy for a 6,000-person mining division
2021
Classroom
Built a state's classroom-observation system for 2,200 mentors, then an AI teacher coach
2025
Lab
Owned a spoken-English practice product; worked on an AI reading assessment's state expansion
2026
Workshop
Building AI simulations for practice, starting with Abhyas Studio
Now
Yale SOM
MBA '28; Education Club first-year lead