Mar 2025 – Present · 1 yr

Jun 2021 – Present · 4 yrs 9 mos

• 1) LLM Training and model optimization
• LLM Finetuning: Worked on finetuning Mistral 70B and LLaMa-2/3 7/8/70B for triplet extraction for building graph RAG on financial Data.
• LLM Model optimization: Optimizing Mistral, Mixtral, and LLaMa and scaling the deployment over large pool of GPUs.
• RAG Pipeline: Building RAG pipleines on various modalities like text, Images, Video and docs for HR bot, Marketting bot, Code Documetnation, Product etc..
• 2) Deep Learning inference over 1000s of live camera
• Architecting platform for large-scale Intelligent Video Analytics
• Model training on multiple GPU, Multiple Node with Mixed Precision. Trained and deployed end-to-end pipeline for Face Detection and Recognition, ANPR, and Intruder Detection for a small object over a long distance.
• I have worked on, Model Training, model profiling, model ensembles, pipeline profiling, component optimization, scaling, and deployment over Kubernetes.
• I have worked with optical flow and tiling-based approaches to make the DNN model compute-friendly.
• Model optimization, Ensemble with different backend/models, Model/application profiling, writing end to end C++/python Deepstream Application, and deployment with Kubernetes orchestration.
• 3) ASR - Indian language support
• Developed standard procedure for ASR and helped customers to develop ASR for multiple Indic languages. Quartznet-15x5 was used, final WER over IITM data was 6.4.
• 4) CUDA: My profile doesn’t demand core CUDA skills but out of my interest, I learned CUDA. This helps in many ways:
• Deepstream postprocessors require custom libraries for each model to do box filtration
• Offloading CPU-intensive processes of the model such as ROI filtering and NMS to GPU.
• Reading existing codes and modifying them for TensorRT custom layers.
• I understand CUDA code at Grid, Block, and Thread levels and can write parallel kernels.

Nov 2019 – Jun 2021 · 1 yr 7 mos

• Unreal Engine | Autonomous Driving | Computational linguistics | Computer vision | GPU Acceleration
• CUDA Computing:
• 1) Efficiency improvement by explicit memory assignment
• 2) Partitioning schemes such as interleaved and block partitioning
• 3) Compressing data using CSR, ELL, and COO techniques, Cuda Parallelism on the compressed data.
• 4) Utilizing Unified memory APIs for faster prototyping
• Nvidia DeepStream
• 1) Writing an application for object detection and event initialization
• 2) Application profiling and fine-tuning
• 3) Large scale deployment of such applications with Kubernetes
• Nvidia Tensorrt
• Model conversion from ONNX, Pytorch, or Tensorflow to FP32, FP16, and INT-8 optimization level
• Multi-Language ASR, and TTS
• 1) Development of models training NEMO API for Quartz-Net and Jasper with lightning-PyTorch
• 2) Optimizing models using Tensorrt and deploying using triton inference server
• 3) Supported language - Hindi, Gujarati, Punjabi
• Deployment at scale
• 1) Deploying dockerized application over Kubernetes
• 2) Grounds up application development that takes advantage of multiple GPU-multiple node environment while deployed.
• Talks:
• 0. GPU Technology Conference session - Building Indic ASR Using NVIDIA NeMo and Deploying Models Using Jarvis : https://gtc21.event.nvidia.com/media/1_kz17xdau
• 1. Nvidia - Accelerated Database Query Using GPU : https://info.nvidia.com/india-accelerated-database-query-reg-page.html?ondemandrgt=yes
• 2. IISER Pune - Nvidia Data Science Ecosystem Tools : https://www.iiserpune.ac.in/events/Workshop+on+Data+Science+Ecosystem+Tools
• 3. Intel and Analytics Vidhya - Representing Language Mathematically : https://www.innoplexus.com/news/the-convergence-of-big-data-and-machine-learning/

Jul 2019 – Nov 2019 · 4 mos

• Unreal Engine | Autonomous Driving | Computational linguistics | Computer vision | GPU Acceleration