Shrota is an AI-powered CATI (computer-assisted telephone interviewing) platform for running phone surveys in India, in Hindi and other Indian languages. See it live at getaicati.com.

A researcher designs the survey in a web builder. Then a voice agent calls each respondent and asks the questions in Hindi. I built all of it on my own, and it is pre-launch right now, running on test calls.

Studies dial contact lists in waves, with call records, live monitoring and per-call tracing. Latency and cost are tracked on every call, and a call carries on with a backup vendor if one fails.

Say an LLM fills a form or a spreadsheet that has to follow a schema, and you want to change one answer or 10 rows out of 10,000. Today the model writes the whole thing again. I wanted to know if it can write only the changed part and splice it back, with the schema still valid.

I ran two experiments on Gemini. Edits took 172x fewer output tokens than writing the whole output again. The code and results are in my research repo.

Built as part of the Co:Rise program in July 2022.

Developed a two-tower candidate generation model and a multi-task re-ranking model for personalized recommendations at scale.

Implemented a contextual multi-armed bandit model with epsilon-greedy policy for exploration-exploitation trade-offs.