Raman Shinde
Senior Lead Engineer, Qualcomm AI Research10+ yrs
I build efficient on-device ML systems — compiler backends, GPU kernels, and the runtime plumbing that lets large models run on small hardware. Deep in TVM, llama.cpp, and Adreno GPU optimization. Previously at Imagination Technologies, Xpanxion, Siemens, and TCS.
Experience
2015 — PresentQualcomm — Senior Lead Engineer · AI Research
- Implemented and upstreamed texture memory support for OpenCL and Vulkan backends in Apache TVM, optimizing data access patterns on Adreno GPUs to achieve a 50% reduction in inference time.
- Implemented Matrix ALU support via cooperative matrix extensions in TVM and llama.cpp Vulkan backends, delivering a 4× speedup in AI workloads and LLM prefill on Adreno GPUs.
- Enabled Q4 quantized inference for Adreno GPUs in the llama.cpp Vulkan backend, expanding on-device LLM compatibility and efficiency.
- Integrated CLML acceleration into GGML and llama.cpp, bypassing generic compute paths to deliver hardware-optimized on-device inference.
- Extended TVM and MLC-LLM with new operators, layers, and runtime features, enabling deployment of LLMs and Diffusion models on edge hardware.
Imagination Technologies — Deep Learning Engineer
- Built a custom AI SDK on Apache TVM to enable high-performance neural network execution across heterogeneous GPU and NNA architectures for edge deployment.
- Implemented LSTM/RNN support in the Neural Compute SDK for PyTorch, TensorFlow, and ONNX, ensuring broader compatibility and accelerated performance.
- Developed custom graph transform passes in Relay IR, improving execution efficiency across diverse workloads.
- Developed quantization toolchains (Post-Training Quantization & QAT) across multiple frameworks, improving model compression and inference speed.
Xpanxion — Data Scientist
- Developed an end-to-end NLP and OCR pipeline for medical document classification and information extraction, leveraging Fonduer, Tesseract, and OpenCV.
- Built reusable AI/ML components including content-based and collaborative recommendation systems and Question-Answering modules.
- Implemented advanced NLP components — NER, knowledge-base-driven QA systems, and sequence-to-sequence translation models.
- Delivered Computer Vision projects spanning object detection, image segmentation, and gesture recognition.
Siemens R&D — Product Development Engineer
- Developed a specialized application within Siemens NX to optimize the design of manufacturing sequences, improving efficiency and precision in assembly line operations.
- Identified and resolved coding issues, contributing to enhanced system performance and reliability across successive NX product releases.
TCS — Software Developer
- Developed monitoring and control applications for NCRA, including real-time issue monitoring, debugging, and GUI modifications based on client requirements.
- Provided support for financial applications (CRD, SRD) for Morgan Stanley, ensuring reliable performance and timely issue resolution.
- Built applications for various clients using Python and C++, contributing to diverse projects and delivering tailored software solutions.
Tech Stack
SKILLSLanguages
CC++PythonCUDAOpenCLVulkan
AI Inference & Compilers
Apache TVMMLC-LLMllama.cppvLLM
Optimization
QuantizationSchedulingParallelismFusion
Frameworks
PyTorchscikit-learnTensorFlowONNX
Deployment & Tools
DockerKubernetesAWSGCPGitMySQLMongoDB
Open Source
CONTRIBUTIONSApache TVMcontributor
Upstreamed Vulkan texture memory support for the Adreno GPU backend, optimizing data access patterns and reducing inference time.
llama.cppcontributor
Implemented Matrix ALU support via cooperative matrix extensions in the Vulkan backend. Enabled Q4 quantized inference and integrated CLML acceleration for Adreno GPUs.
Solo Ventures
SIDE PROJECTS
IngreliShutdown2025
A mobile app that helps families make informed food choices by scanning product labels for simple ingredient analysis, safety alerts, and healthier alternatives.
- Allows users to create personalized health profiles for their entire family to track unique dietary goals and restrictions.
- Instantly alerts users to harmful ingredients like allergens and carcinogens using a powerful OCR label scanner.
- Recommends healthier product alternatives tailored to individual dietary needs and preferences.
MedyzerShutdown2022 — 2023
A SaaS platform designed to help doctors digitize their clinic workflow, from patient appointments to digital prescriptions.
- Enabled doctors to manage appointments, patient records, and prescriptions across multiple clinics.
- Provided patients with QR-code appointment booking and access to their digital health records.
- Successfully onboarded practicing doctors and generated over 1,000 digital prescriptions.
Blogs
2 POSTSFEB 21, 2025
Imports in Solidity
Understanding imports in Solidity: local imports, importing directly from GitHub, and handling dependencies in Brownie.
blockchainsolidityethereumsmart contracts
FEB 20, 2025
Exploring the World of AI Compilers: Powering AI Compute and Inference
From NVIDIA TensorRT to OpenAI Triton, discover the AI compilers revolutionizing performance across GPUs, CPUs, and edge devices.
compilersTVMtritonIREEONNX