I am an algorithmic and systems-focused Computer Vision Engineer with a holistic grasp of the vision stack. My expertise bridges the gap between classical computer vision techniques (image processing, geometric reasoning, and tracking), deep learning architectures, and high-performance production pipelines: from custom training loops and multi-process system design to model optimization and edge deployment (TensorRT, Qualcomm AI Hub, ONNX, TFLite).

• IEEE LPCVC 2026 (CVPR Workshop) – Track 2: Ranked 8th/38 globally in Video Action Recognition. Optimized R2+1D for Qualcomm Dragonwing IQ-9075 under a strict <34ms latency budget, overcoming severe class imbalance and label noise through rigorous data auditing.
• IEEE LPCVC 2026 (CVPR Workshop) – Track 1: Ranked 12th/56 globally in Open-World Retrieval. Optimized MobileCLIP/ViT-B/16 for Qualcomm XR2 Gen 2, engineering a custom model wrapper to resolve tokenizer discrepancies (Causal vs. Bidirectional) for on-device inference.
• High-Performance Systems Architecture: Redesigned a monolithic, single-process 30 FPS dual-camera pipeline into a 4-process parallel architecture using Zero-Copy IPC (multiprocessing.shared_memory) and OpenCV-CUDA. Achieved over 3x throughput (30 → 100+ FPS per camera stream) by resolving GUI latency and CPU-bound preprocessing bottlenecks.

• Systems Architecture & Optimization: High-Performance Multiprocessing, Zero-Copy IPC (Shared Memory), GPU Acceleration (OpenCV-CUDA), TensorRT, Qualcomm AI Hub (QNN), ONNX, INT8 Quantization, Structured Pruning.
• Deep Learning & DL-based CV: Object Detection (YOLO v5-v11), Segmentation (U-Net), Tracking (ByteTrack, TransReID), Pose Estimation, Vision Transformers (ViT/VLMs), ReID, Video Action Recognition (Temporal Modeling), Multi-modal Retrieval (CLIP, MobileCLIP).
• Classical & Algorithmic CV: Image Processing (Morphology, CLAHE, Canny, Contours), Geometric Reasoning (Zhang's Calibration, Homography, ArUco, scipy.optimize), Tracking (Kalman Filtering, Optical Flow, Background Subtraction).
• Languages & Tools: Python (Expert), PyTorch, TensorFlow, C++ (CUDA Certified), Netron Model Surgery, Reverse Engineering, Model Benchmarking (clip_benchmark).

• Swift Vision Pvt. Ltd. | Computer Vision Intern → Computer Vision Engineer → Computer Vision Research Engineer | Apr 2025 - Jan 2026 (Promoted to R&D ownership within 4 months based on execution and technical leadership).
• Folio3 Pvt. Ltd. | AI Intern | July 2023 - Sep 2023

• Role: Lead Systems Architect (Swift Vision)
• Challenge: Client's monolithic 30 FPS dual-camera sports analytics system was bottlenecked by sequential CPU-bound processing and OpenCV GUI latency. The client initially misdiagnosed the issue as a model inference problem.
• Solution:
  • Diagnosed the architectural flaw and argued against the TensorRT misdiagnosis.
  • Redesigned the system into a 4-process parallel architecture: two camera logic processes (GPU preprocessing + YOLO inference), an update/state loop, and a main GUI process.
  • Implemented Zero-Copy IPC using multiprocessing.shared_memory to eliminate data copying overhead.
  • Offloaded all image preprocessing (resize, warp, blur, dilation) to OpenCV-CUDA.
• Result: Achieved over 3x throughput (30 → 100+ FPS per stream, with peaks near 120) and 120 FPS GUI rendering. Received direct praise from the client for the architectural solution.

• Result: Ranked 8th/38 globally.
• Challenge: 92-class fitness video dataset with severe class imbalance (some classes <150 examples) and significant label noise.
• Solution & Approach: Modified the training pipeline to log per-class train/val accuracy and individual misclassifications, manually auditing 4,500+ problematic instances. Designed 7 separate augmentation and cropping strategies based on class-specific confusion analysis. Diagnosed a Qualcomm hardware compiler tiling failure and pivoted to a 16-frame architecture to meet a <34ms latency budget.
• Outcome: Improved accuracy from 91.44% to 93.24% through rigorous data curation and training rule refinement.

• TensorRT validation fix (PR #21592, merged) — Root-caused a yolo val failure on TensorRT .engine models (the validator read a batch_size attribute that AutoBackend never set) and worked through maintainer review; the merged fix simplified the validator's batch-size logic.
• Structured Pruning Engine (PR #21977, docs PR #22438) — Native PyTorch structured pruning for YOLOv8 detection models.
  • Global ratio or per-layer YAML configuration; norm-based channel importance; mask propagation through Conv, BatchNorm, Bottleneck, C2f, SPPF, Concat and both Detect-head towers; group-aware handling of grouped/depthwise convolutions (the reason torch.nn.utils.prune wasn't enough).
  • ~1,000 lines of pruning code plus a 1,400-line test suite (43 test functions) covering the prune → save → load → retrain → ONNX export round trip.
  • Functional demo (YOLOv8s, COCO128, ONNX): 22.6 → 11.6 MB (~49% smaller), 17.1 → 12.0 ms (~30% faster). A demonstration that the pipeline works end to end, not an accuracy benchmark; retraining is needed to recover accuracy.
  • Status: scoped with maintainer guidance, CI green and branch up to date; both PRs were closed by the stale bot before code review. Code is on my fork: feature/prune-functionality and docs/pruning.

Reimplemented from the original papers, first principles.

• Transformer (Attention Is All You Need) — Complete preprocessing and training pipeline, shared source-target vocabulary, byte-pair encoding, three-way weight tying, padding and causal masking, full encoder-decoder architecture.
• U-Net (Biomedical Segmentation) — Pixel-wise weight maps and elastic deformation, optimized for TPU-accelerated training using a TFRecord-based data pipeline.

• Real-Time Liquid Quality Inspection — Automated inspection pipeline for transparent bottles on a conveyor using backlit imaging; morphological segmentation, CLAHE, Canny edge detection, contour analysis, and Kalman filter tracking.
• Clinical Caries Detection — Customized U-Net for dental radiography segmentation, high-precision boundary detection.
• Traffic Analytics Pipeline — Automated traffic analysis using optical flow, thresholding, and multi-object tracking (MOT) to monitor vehicle flow and density.
• Deep Probabilistic Generative Models — VAEs for facial generation and RealNVP (Normalizing Flows) for high-dimensional image synthesis on LSUN.
• Bayesian CNN (Uncertainty Quantification) — Captures aleatoric and epistemic uncertainty in digit classification, beyond point-estimate predictions.
• Bayesian Earthquake Forecasting — Bayesian AR(3) model in R for global seismic activity: model order selection, posterior inference, prior sensitivity, mixture AR model comparison..

A collection of visual demos from early project work showcasing classical CV, tracking, and basic detection pipelines.

• View the full project demo playlist on YouTube (Includes: Water Quality Inspection, Vehicle Detection, Pose Estimation, Pedestrian Tracking, Multi-Car Tracking, Liquid Fill Estimation, and more).

• Foundational Exposure: Studied early chapters of Hartley & Zisserman's Multiple View Geometry to understand the mathematical principles behind production geometry tasks.
• Applied Practice: Translated these geometric foundations into working code, including Zhang's planar calibration and marker-aware ArUco rectification using scipy.optimize.
• Current Learning Focus: Currently building depth in Kalman Filtering and C++ for high-performance geometric vision.
• Upcoming Study Plan: Next steps include 3D vision libraries (PCL) and advanced coursework (CS231A) to deepen 3D perception expertise.

• BS Computer Science | IBA Karachi (Best Paper Nomination, ICETST 2022)

• Email: syedhamza097@gmail.com
• LinkedIn: linkedin.com/in/syedhamzamohiuddin
• Kaggle: kaggle.com/hamzamohiuddin
• Medium: medium.com/@syedhamza097