Data scientist working on speech and language — ASR, TTS, and LLMs.
Introduction
I'm a Data Scientist at Meesho working on the speech stack — Hindi Automatic Speech Recognition including code-mixed speech, Text-to-Speech for Malayalam, Telugu and Tamil, and Large Language Models fine-tuned for narrow jobs like evaluation. I came to it sideways: two years building insurance software, then a master's at IISc spent teaching small models to hear crackles and wheezes in stethoscope recordings. Most of what I've done since is a version of that same problem — audio in, a decision out, and an honest measure of whether it worked.
Hindi ASR on real call audio, including code-mixed speech where people switch to English mid-sentence — served at 30 RPS with p90 100 ms latency.
Text-to-Speech for Indian languages that have far less training audio than English. Output quality assessed by human auditors.
Fine-tuning Large Language Models for specific, testable jobs — like user simulation, so prompts can be evaluated in a loop instead of against live traffic.
Spectral features and small architectures. My thesis found compact CNNs beating wav2vec 2.0 on lung sound classification — bigger isn't automatically better.
Where I've worked
Where I studied
What I've built
Detecting crackles and wheezes in stethoscope recordings, and classifying the underlying condition. Compact CNNs over mel-spectrogram and MFCC features beat every wav2vec 2.0 variant tested — 78.7% on four-class detection.
Hindi ASR on a fine-tuned IndicConformer — first on Devanagari script, then extended to code-mixed speech where users switch to English mid-sentence. Served through TensorRT (FP32) at 30 RPS, p90 100 ms / p99 200 ms, for 60% less cost.
XTTS fine-tuned for Malayalam, Telugu, and Tamil — three languages with far less training audio than Hindi or English. Output quality assessed by human auditors.
Ranks sellers by conversion affinity and scores call slots hour by hour, so calling effort goes where it pays, with an explore–exploit policy for cold-start sellers. Cut outbound calling cost by 50%.
A fine-tuned Llama 3.2 3B Instruct that plays the user side of a call, so prompts and conversation designs can be evaluated in a tight loop instead of against live traffic.
Published work
V. Mehra, D. Pandey
International Journal of Software Innovation (IJSI), 2022
V. Mehra, D. Pandey, A. Rastogi, A. Singh, H. P. Singh
Recent Patents on Engineering, 2021
Get in touch
Happy to talk about signal processing, NLP, or anything I've written about here.