Headline

Machine Learning Engineer @TCS

About

Engineer focused on LLM optimization and edge AI deployment, currently working as consultant through TCS with Qualcomm's team on NPU model certification and deployment infrastructure. At Qualcomm, I work on certifying AI models for NPU execution, building automation pipelines for model validation, and tracking AI workload deployments across production environments. At TCS, I design distributed inferencing architectures for enterprise clients, having optimized LLM inference throughput by 40% and reduced VRAM consumption by 60% across large-scale deployments. My core interest is in making AI models run efficiently whether that's through quantization (AWQ, GGUF, fp8), distributed systems (Ray, vLLM), or hardware-aware deployment on GPUs and NPUs. I've worked across the stack: from benchmarking 20+ LLMs and evaluating hardware performance on NVIDIA and Qualcomm silicon, to building RAG systems and computer vision models earlier in my career. Always open to connecting with people working on inference infrastructure, edge AI, and ML systems.

Location

Bengaluru

AI stack

claude

langchain

langgraph

langsmith

mistral

ollama

Work experience

Engineer - I (Temp) at Qualcomm, Jan 2026 - Present · 8 mos, Bengaluru, Karnataka, India, - Certifying 12+ AI models for NPU execution, validating Execution Provider (EP) compatibility and ensuring compliance with Qualcomm hardware constraints - Automating the end-to-end model certification pipeline using Axiom, reducing manual validation overhead across the team

Machine Learning Engineer at Tata Consultancy Services, Feb 2025 - Present · 1 yr 7 mos, Bengaluru, Karnataka, India, - Designing distributed inferencing architecture for enterprise-scale RFP cycles - Engineered multi-node GPU clustering across heterogeneous systems, optimizing inter-node communication for client latency requirements - Benchmarked 20+ LLMs across prompt templates and context windows to establish performance grading metrics for code summarization - Optimized LLM inference throughput by 40% using tensor parallelism, quantization (AWQ, GGUF), and Ray - Reduced VRAM consumption by 60% for Llama 3 70B deployment via vLLM with CPU offloading and paged attention - Benchmarked Llama 3.2 1B across quantization techniques (AWQ, bitsandbytes, fp8) and evaluated WinML performance on NVIDIA and Qualcomm hardware

System Engineer at Tata Consultancy Services, Jan 2025 - Feb 2025 · 2 mos, Bengaluru, Karnataka, India

Undergraduate Researcher at Cognitive Computing Research Centre (CCRC), May 2023 - Jun 2024 · 1 yr 2 mos, India, - Developed and implemented classification models, for pothole classification on a self-curated dataset. - Achieved notable F1-scores: 63.41 with MobileNetV2, 61.47 with VGG16, and 61.25 with ResNet-18, showcasing successful pothole classification capabilities. - Utilized a diverse technology stack inclusive of Hugging Face’s Transformer models for natural language processing tasks such as text summarization and semantic analysis.

Education

Bachelor of Science, Computer Engineering, Federal Institute of Science and Technology

Bachelor of Technology - BTech, Computer Science, APJ Abdul Kalam Technological University