Samunder Singh

Design Engineer at C-DAC Bangalore

GPU Design Engineer / AI/ML Engineer / Open Source Contributor

Building LLMs for the edge, Vision AI, and the software stack behind custom GPU hardware.

about

I'm an AI/ML Design Engineer working at the intersection of LLMs, distributed systems, GPU computing, and AI infrastructure. Currently at C-DAC Bengaluru, I build 1B–7B small language models with a focus on distributed training, inference optimization, quantization, and deployment. I also work on the software stack for custom GPU hardware — OpenCL kernels, RISC-V, POCL, compiler/runtime development — and enabling PyTorch-like AI workloads on custom accelerators.

Previously at Lincode I built and deployed production ML/CV systems using PyTorch, YOLO, DETR, CLIP, TensorRT, torch.compile, and distributed training. I'm interested in LLM systems, GPU programming, compilers, AI accelerators, and hardware–software co-design. I also contribute to open source, including Ivy at Unify.

B.Tech in Electrical, Electronics and Communications Engineering · Tezpur University, Assam, India · 2020 - 2024

experience

C-DAC — Design Engineer
July 2026 - Present
Bengaluru, India
- Building end-to-end infrastructure for distributed SLM training and inference — training pipelines, inference optimization, and quantization for in-house 1B–7B parameter models.
- Developing the software stack for a custom GPU accelerator at the compiler/runtime level with OpenCL kernels, RISC-V POC, and POCL.
- Enabling PyTorch-like AI frameworks and LLM inference on custom hardware, bridging low-level kernels to high-level workloads.
Lincode — Associate Data Scientist
September 2024 - July 2026
Bengaluru, India
- Built a synthetic data pipeline with a structured defect library — generation, augmentation, and dataset versioning — for rare-class imbalance in industrial inspection.
- Designed a self-supervised auto-labeling algorithm and a CLIP-based unsupervised pipeline for anomaly detection and data visualization.
- Trained and deployed YOLOv8, DETR, and CLIP for defect detection; optimized inference with TensorRT and torch.compile for edge and cloud serving.
- Scaled distributed training across multi-GPU and multi-host clusters using PyTorch DDP.
- Led end-to-end integration of segmentation and OCR (PaddleOCR) into real-time inspection workflows.
- Integrated multi-bucket cloud storage (AWS S3, DigitalOcean Spaces, Storj) for datasets and model artifacts.
- Built an MCP server and custom AI tools for LLM-driven automation of internal workflows.
- Enhanced a Django-MongoDB backend for scalable ML product deployment; used K-Means, DBSCAN, Grad-CAM, and SHAP for discovery and explanations.
Lincode — Data Science Intern
May 2024 - September 2024
United States (Remote)
- Developed and integrated object detection models (YOLO, DETR) into the product pipeline, improving real-time inspection accuracy.
- Built image augmentation pipelines to improve generalization across manufacturing scenarios.
- Contributed to the Django backend — bug fixes, API performance, and ML model integration.
- Helped ship a segmentation and OCR pipeline with PaddleOCR into production.
Unify — Top Contributor
February 2024 - May 2024
Remote
- Collaborated on test-case fixes and critical Ivy tasks, including deep dives into transpiler and compiler code and PR review.
- Built a chatbot to compare and rank LLMs with a global leaderboard — users query two models at once and vote on the winner.
- Worked on complex dtypes across TensorFlow, JAX, and PaddlePaddle backends, including SHLO/XLA improvements.

projects

SLX
Deep Learning
A high-performance deep learning framework on PyTorch with custom shader-based kernels through Slang, for efficient tensor operations on GPU.
Slang PyTorch Python CUDA
Shruti-Drishti
Computer Vision
Real-time Indian Sign Language recognition with Mediapipe, CNN, and LSTM at 90%+ accuracy. Won an award at a Hasgeek hackathon.
Python MediaPipe TensorFlow Flask
Edulens
Computer Vision / NLP / LLM
FastAPI video analysis platform using ML for figure detection, OCR, and speech-to-text, with Google OAuth and Vectara.
FastAPI Celery Google OAuth VectorDB
LLM Playground
NLP / LLM
Chatbot Arena for benchmarking LLMs in the wild — enter a prompt and compare two models side by side.
Docker GCP Streamlit LLMs
YOLO Agent
Computer Vision / LLM
LangChain agent that talks to Ultralytics YOLO so you can train, validate, export, and deploy detection models in natural language.
YOLO LangChain Python OpenAI
Sign Language Detection
Computer Vision
Translated real-time sign gestures into text using holistic landmark tracking with MediaPipe and TensorFlow.
Python MediaPipe TensorFlow Flask
Travel Chatbot with RAG
NLP / LLM
LLM chatbot that fetches live weather and search results with LangChain and OpenAI.
LangChain OpenAI Streamlit RAG
YOLOv7 Custom Training
Computer Vision
Custom object detection on an aquarium dataset with visual logging and monitoring.
YOLOv7 PyTorch W&B OpenCV

skills

LLM Systems: Distributed Training · Quantization · SLMs (1B–7B) · TensorRT · torch.compile · PyTorch DDP
GPU & Hardware: OpenCL · RISC-V · POCL · CUDA · Compiler / Runtime
Languages: Python · C++ · SQL · Hindi · English
ML & Vision: PyTorch · YOLOv8 · DETR · CLIP · PaddleOCR · OpenCV · MediaPipe · JAX · TensorFlow
Backend & Cloud: Django · FastAPI · MongoDB · Docker · AWS S3 · MCP · Git

certifications

Generative AI with Large Language Models — AWS, deeplearning.ai
Neural Networks and Deep Learning — deeplearning.ai
Supervised Machine Learning: Regression and Classification — deeplearning.ai
Structuring Machine Learning Projects — deeplearning.ai
Foundations: Data, Data, Everywhere — Google

writing

recommendations

Samunder has done great work during his time as a top contributor working on Ivy. His enthusiasm was infectious to those around him, and he learned a lot during the role, picking up new skills very quickly, with a drive to get things done. I would recommend him for any role involving Machine Learning infrastructure, and especially those using a lot of Python!
Daniel Lenton — CEO at Unify | YC W23 · May 6, 2024