Muhammad S J is an AI Engineer with several years of experience in artificial intelligence, specializing in machine learning, deep learning, and computer vision for large-scale applications. Proficient in Python, C, and C++, Muhammad utilizes frameworks like TensorFlow, PyTorch, OpenCV, and MMDetection to build high-performance AI models. He constructs end-to-end machine learning pipelines, optimizing and deploying models across environments such as ONNX and edge devices. His expertise includes CUDA optimization,… Show more

1. Algorithm Engineer

   Software MotionJul 2024 — Nov 2025

   • Fine-tuned and trained the FCOS3D version-3 model on the nuScenes dataset and custom lidar-camera inputs by Nexdata.
   • Curated the training pipeline by integrating custom transforms, applying domain adaptation strategies and optimizing hyperparameters I was able to increase the FCOS3D object detection capability by 4%.
   • Fine-tuned FCOS3D was converted to ONNX for deployment on x86 processors by updating the SDK from an outdated version.
   • Ported the optimized model to the Huawei MDC510 board by converting ONNX → OM format and rewriting pre/post-processing in C++ for the updated version of fcos3d.
   • Executed end-to-end deployment by logging into the board from x86, uploading binaries, dataset and model, and successfully running inference on target hardware.
   • Conducted data validation of our custom data via calibrating camera intrinsic, extrinsic and distortion coefficients by mapping 2D and 3D bounding boxes onto ground truths, distorted and undistorted images and PCD’s, to ensure accurate training data for our object detection pipeline.
   • Improved CondLaneNet by modifying postprocessor logic, enhancing Breakpoint Detection and Mask Proposal processing, Improved Vector Field Processing, achieving a 5% accuracy gain in lane detection.
   • Integrated CondLaneNet into multi-head model alongside FCOS3D (3D object detection) and traffic light recognition(TLR), enabling a unified perception pipeline for autonomous driving. Experienced in Rosbag parsing, for extracting data and performing validation tasks.
   • Analyzed issues and collaborated with calibration, perception fusion and mapping teams to find effective solutions.
   • Worked with Parking Lane Detection and delivered detailed presentation about its architecture, pre/post processing codes and accuracy scores that we obtained by training from scratch onto the dataset available online.

   ONNXC++pythonmmdetection3dmmdetection
2. Senior AI Engineer

   edgec.ioNov 2025 — Jan 2026

   • Led a small ML Team and developed computer vision models on hyperspectral satellite imagery to detect and map energy infrastructure including oil terminals, refineries, rigs, solar parks, and wind farms.
   • Built end-to-end ML pipelines for data preprocessing, model training, evaluation, and scalable inference on large geospatial datasets.
   • Designed and implemented a Retrieval-Augmented Generation (RAG) chatbot, owning the full backend and system architecture.
   • Created a financial knowledge base (~85k documents) by extracting structured financial tables from unstructured data and embedding them for semantic search.

   pythonRAGVector data basesLLMVLM'sVAPIseasme.ai
3. Computer Vision Engineer

   CCRIPT AgencyJan 2026 — Present

   • Developing an end-to-end estimator that ingests architectural, structural, and civil plan sets to autonomously extract material quantities (e.g., concrete volumes, wall lengths) without manual intervention.
   • Engineering a pipeline that converts raw extraction data into professional, priced Bid Book PDFs and editable Excel takeoffs, ensuring outputs are fully traceable and compliant with standard civil bid submission formats.

   Construction EstimatorsPythonReactNode.js

1. Bachelor of Science in Artificial Intelligence

   Ghulam Ishaq Khan Institute of Engineering Sciences and TechnologyTopi2020 — 2024