Looks like you have JavaScript disabled. For the full Twine experience, you will need to re-enable it.

AI software engineer specializing in deep learning, multimodal computer vision, and CUDA/C++ performance optimization. Built end-to-end real-world systems for recognition, tracking, OCR, analytics, and edge deployment—delivering strong metrics and ahead-of-schedule execution. Focused on scalable ML infrastructure, model evaluation/monitoring, and practical deployment (Jetson/TensorRT/ONNX). Active research across vision-LLMs, spatial reasoning, and agentic LLM frameworks for accelerating high-performance CUDA engineering.…AI software engineer specializing in deep learning, multimodal computer vision, and CUDA/C++ performance optimization. Built end-to-end real-world systems for recognition, tracking, OCR, analytics, and edge deployment—delivering strong metrics and ahead-of-schedule execution. Focused on scalable ML infrastructure, model evaluation/monitoring, and practical deployment (Jetson/TensorRT/ONNX). Active research across vision-LLMs, spatial reasoning, and agentic LLM frameworks for accelerating high-performance CUDA engineering.

Siddharth Agrawal

AI Engineer, Software Engineer, Data Scientist, +4





AI software engineer specializing in deep learning, multimodal computer vision, and CUDA/C++ performance optimization. Built end-to-end real-world systems for recognition, tracking, OCR, analytics, and edge deployment—delivering strong metrics and ahead-of-schedule execution. Focused on scalable ML infrastructure, model evaluation/monitoring, and practical deployment (Jetson/TensorRT/ONNX). Active research across vision-LLMs, spatial reasoning, and agentic LLM frameworks for accelerating high-performance CUDA engineering.…AI software engineer specializing in deep learning, multimodal computer vision, and CUDA/C++ performance optimization. Built end-to-end real-world systems for recognition, tracking, OCR, analytics, and edge deployment—delivering strong metrics and ahead-of-schedule execution. Focused on scalable ML infrastructure, model evaluation/monitoring, and practical deployment (Jetson/TensorRT/ONNX). Active research across vision-LLMs, spatial reasoning, and agentic LLM frameworks for accelerating high-performance CUDA engineering.

Available to hire

AI software engineer specializing in deep learning, multimodal computer vision, and CUDA/C++ performance optimization. Built end-to-end real-world systems for recognition, tracking, OCR, analytics, and edge deployment—delivering strong metrics and ahead-of-schedule execution.

Focused on scalable ML infrastructure, model evaluation/monitoring, and practical deployment (Jetson/TensorRT/ONNX). Active research across vision-LLMs, spatial reasoning, and agentic LLM frameworks for accelerating high-performance CUDA engineering.

Skills

Experience Level

Expert

Expert

Expert

Expert

Expert

Expert

Expert

Expert

Expert

Expert

Expert

Expert

Expert

Intermediate

Intermediate

Intermediate

Language

English

Advanced

Hindi

Advanced

Work Experience

Research Automated LLM Agentic Framework for CUDA Engineering at New York University

August 1, 2025 - December 1, 2025

Engineered a multi-agent LLM framework to automate high-performance C++/CUDA code generation using strategic tree search and hardware profiling with NVIDIA Nsight Compute. Achieved 95% correctness and 1.28x average speedup on KernelBench, surpassing state-of-the-art baselines.

Enhancing Spatial Reasoning in Vision-LLMs at New York University

January 1, 2025 - May 1, 2025

Developed conversational chain-of-thought and structured prompt pipelines to improve spatial reasoning in multimodal VLMs by ~8–24% on spatial reasoning benchmarks. Implemented reinforcement learning formulations, improving Pass@1 accuracy by up to 23% on some tasks.

AI Software Engineer (R&D) at The Main Branch

January 1, 2024 - August 1, 2025

Engineered end-to-end computer-vision recognition and tracking system for cattle management using CCTV, achieving 82.7% MOTA and 0.65 ARI by distilling SAM-2 into FastSAM and optimizing CNNs for ~6,000% faster training and real-time visual search. Expanded pipeline to include disease prediction, weight estimation using LiDAR/depth maps, action recognition, and feed/hydration analytics. Built scalable ML infrastructure with vector search and time-series partitioning to manage 100M+ measurements with SQL-aggregated stakeholder dashboards. Performed low-level NVIDIA Jetson/CUDA optimizations (custom kernels, pruning/quantization, TensorRT) for major speedups and faster inference. Led mentoring of interns, synthetic data generation, model adaptation, and CVAT/Roboflow annotation workflows. Delivered an OCR system with 98.3% accuracy integrating LLM/NLP for multimodal summarization.

ML & Software Engineering (Independent Contractor) at Office of State Minister, Ahmedabad

February 1, 2023 - August 31, 2023

Built AI models using CCTV footage for law enforcement, traffic flow, and license plate recognition. Developed tracking for traffic violations including seatbelt usage, phone usage, and red-light running to support automatic fining workflows.

ML & Software Engineering (Independent Contractor) at Office of State Minister

February 1, 2023 - August 1, 2023

Built AI models on CCTV footage for law enforcement, traffic flow, and license plate recognition. Developed a tracking system for traffic violations including seatbelt usage, phone usage, and running red lights to support automatic fining workflows.

Undergraduate Thesis – Global License Plate Dataset at Ahmedabad University

November 1, 2022 - March 1, 2024

Created a 5M+ annotated image dataset from 74 countries for license plate detection and recognition using semi-supervised learning and confident learning for semi-automated annotation.

AI Full-Stack Engineer Intern at Sensegood Instruments

May 1, 2022 - February 1, 2023

Used YOLOv8 and synthetic data generation to create datasets for wheat and rice grain segmentation, detecting coloration defects, broken grains, and impurities (dirt/pebbles/grass/straw), achieving a 98.5% F1-score.

Handwritten License Plate Recognition

December 1, 2021 - November 1, 2022

Optimized PARSeq Vision Transformer for handwritten Indian license plates using aspect ratio augmentation and vertical image-text concatenation, plus synthetic data generation; improved accuracy from 58.96% to 95.82%.

Research Assistant – Digitizing and Topic Modeling of US Presidential Directives

May 1, 2021 - October 1, 2021

Performed OCR digitization and exploratory data analysis using transformers (BERT, RoBERTa, XLNet, Longformer) to visualize distributions, clustering, and topic modeling; created visualization/mining workflows using similarity matrices, intertopic distance maps, hierarchical clustering, and dynamic topic modeling.