About me

Hello! I am Romit. I love building data-driven tools and products. I have (Current year - 2018) years of work experience in arranging matrices.

A bit about my experience

• ML engineer at Sarvam.ai, [February 2025-Present]
  • Currently working on pre-training vision encoders for sovereign vision language models.
  • Led post-training quantization and inference optimization of Sarvam-M, a 24B LLM - Section 4 in the blog.
  • Built a highly fault-tolerant, scalable central store for pre-training datasets using WebDatasets.
  • Worked with TensorRT LLM, vLLM, WebDatasets, and K8s.
• AI engineer at Meraki Labs, [December 2023-January 2025]
  • Pre-trained small audio language models. This model aligned audio and text tokens into the same space, enabling TTS, ASR, and text-to-voice continuation from a single model.
  • Spent time understanding the audio domain and how its modeling works.
  • Developed custom GPU kernels in Triton to accelerate inferencing and researched inference bottlenecks in the audio domain.
  • Worked with Triton, CUDA C, and PyTorch.
• Senior Data Scientist at epifi, [November 2020-October 2023]
  • Built customer-support call categorization pipelines using Whisper, Mistral 7B, and fine-tuned BERT.
  • Trained and deployed real-time and batch fraud-detection models using computer vision and tree-based methods.
  • Built infrastructure for real-time and batched feature engineering and optimized model deployment.
  • Worked with Tensorflow, Docker, PySpark, Airflow, EFK stack, and the usual AWS stack.
• Technical Lead at Postman, [January 2018-November 2020]
  • Founding member of the data team; built an analytics platform supporting over 1,000 production dashboards.
  • Architected data testing frameworks and data warehouse table design to power business analytics.
  • Worked with SQL, Redshift, Looker, DBT, and PowerBI.
• I was on the team that built the project Panoptic.

Places where I have talked about my work experience

• Gave a talk on efficient inferencing at AI Inference Meetup with NVIDIA.
• Gave a demo on audio language models at AI Tinkerers Bengaluru chapter.
• Gave a workshop at Fifth Elephant conference on writing efficient Triton GPU Kernels.
• Speaker at 2022 AWS ML Fridays series for fraud detection using Amazon SageMaker.

A bit about what I love doing

• Tech I love working with - Python, Triton, PyTorch, SQL, Docker, FastAPI.
• Areas I am interested in:
  • Systems around LLMs
  • Applied deep learning, especially computer vision
  • Human in-loop decisions
• Reading books, playing cricket, and petting cats.
• Obsessed with calendars, to-dos, tracking every little thing, and shiny productivity apps.