Donya Khaledyan

Find Me

• LinkedIn
• Google Scholar
• GitHub

Professional Summary

AI and Computer Vision Scientist with 6+ years of experience specializing in deep learning, computational imaging, hardware design, and model optimization (hardware and software). Expertise in image processing, signal processing, and medical imaging workflows, with hands-on experience in FPGA and hardware implementation, and real-time algorithm development. Proven track record of building scalable AI systems and integrating ML/GenAI into clinical pipelines. Strong background in multimodal deep learning, generative AI (GANs, VAEs, Diffusion Models), and optimization techniques.

Key Skills

• Programming: Python, MATLAB, C, C++, VHDL, Verilog, Assembly.
• Data Science & HPC: TensorFlow, PyTorch, Pandas, CUDA, cuDNN, Slurm.
• Deep Learning & AI: CNNs, NLP (RNN, LSTM, Transformers), LLMs (Llama, BERT, GPT), Generative AI (GANs, VAEs, Transformers, Diffusions), Multimodal AI.
• Machine Learning: KNN, SVM, Decision Trees, Random Forest, Bayesian Networks.
• Computer Vision: Scikit-Image, OpenCV, MATLAB Image Toolboxes, 3D vision models (3D UNet, V-Net, SLAM, stereo matching, NeRFs).
• Optimization: Quantization, knowledge distillation, pruning, CUDA kernel optimization, ONNX, TensorRT, OpenVINO.
• Hardware Acceleration: FPGA implementation, real-time hardware systems, edge computing.

Skills Proficiency

Professional Experience

Specialized HW and ML Engineer Intern — Amazon Lab126 | Sunnyvale, CA | 06/2025 – 09/2025

• Goal-driven, resource-constrained design: Proposed and delivered color-mixing/display-enhancement solutions optimized for the limited compute/memory/power budget on Kindle Scribe, prioritizing small, fast, and reliable models over heavyweight DL.
• Physics-based data engine: Generated a large, diverse synthetic dataset using physically inspired mixing, plus measured swatches, to cover pigments, paper textures, and stroke behaviors across the device gamut.
• Model selection under scaling: Benchmarked compact learners end-to-end:
  • Polynomial regression as a strong classical baseline, but observed instability and poor scalability in feature dimensionality & dataset size (fit time, numerical conditioning).
  • Small MLPs ( narrow width) emerged as the best accuracy–latency–memory trade-off, capturing nonlinearity without heavy runtime cost.
• Full developement & deployment: PyTorch → C++ inference (LibTorch/OpenVINO paths), with static-shape configs, operator fusion, thread pooling, and memory reuse.
• Device integration: Embedded model into the Kindle Scribe pipeline with LUT export fallback, color-profile hooks, and guardrails for out-of-gamut inputs; exposed tunables for stroke sharpness and saturation ramp.
• User study impact: Post-integration UX tests showed clear user preference for the new models vs. legacy blending (higher perceived realism and faster color convergence), supporting rollout decisions.
• Key Outcomes: Production-ready, compact MLP models; physics-grounded data generation; complete PyTorch→C++ deployment; measurable UX wins on device.

Research Scientist Intern -- Meta Reality Labs | Redmond, WA | 05/2024 – 11/2024

• Developed deep-learning models in PyTorch and Keras for polarization computational imaging in AR/MR/VR applications, improving reconstruction accuracy by 50% compared to SOTA baselines.
• On raw and sparse RGB-polarization (DoFP) data, built a calibrated ISP (flat-field, cross-talk, Stokes→AoLP/DoLP) and optimized GAN/ VAE/ ViT/ diffusion models for joint demosaicing–depolarization, inpainting, and super-resolution, yielding higher-fidelity RGB-polar reconstructions with fewer artifacts compared to all the baselines.
• 3D rendering and digital twins using NeRF and 3D Gaussian Splatting, integrating polarization sensors/images (AoLP/DoLP) to enhance depth, surface normals, and material estimation for VR.
• Key Skills: NeRF, 3D Gaussian Splatting, Polarization Imaging, GANs, VAEs, ViTs, Diffusion Models, Camera & Sensor Systems, 3D Reconstruction, Depth Estimation, HPC, Docker.

AI Researcher & Team Manager -- TECVICO | Vancouver, Canada | 05/2020 – 11/2021

• Led a team of AI researchers to develop medical imaging models.
• Designed and deployed CNN and transformer-based frameworks in Keras for healthcare anomaly detection.
  • Head & Neck Cancer (MRI + CT): Built a multi-modal tumor segmentation and lymph-node detection pipeline with deformable registration between MRI–CT; achieved oncologist-validated Dice improvements and integrated RT-structure exports (DICOM-RT) to assist radiation therapy planning and target delineation in clinical workflows.
  • Breast Imaging (Photon-Counting CT & Conventional CT): Developed lesion detection and tissue-density quantification using RetinaNet/YOLOv8 for candidate generation and ViT-based classifiers for BI-RADS-style triage; supported second-read decision support and dose/protocol optimization in industry PACS pipelines.
  • Anomaly Detection @ Scale: Deployed self-/weakly-supervised pretraining (SimCLR/MoCo, masked autoencoders) on mixed MRI/CT cohorts to flag out-of-distribution scans (artifacts, motion, protocol mismatch); used for worklist triage and automated QA in hospital imaging networks.
  • Survival Prediction: Survival Prediction in head and neck cancer based on employing Deep Learning methods such as convolutional neural networks (CNN).
• Key Skills: Team Management, System Design, Medical Imaging, Anomaly Detection.

PhD Research Assistant -- University of Rochester | 01/2022 – 11/2025

• Developed multimodal DL models in PyTorch for enhanced image-to-image translation in OCE (Optical Coherence Elastography), conditioning on acquisition metadata (e.g., scan depth, sampling rate, probe settings, tissue region, case age, awake/ sleep condition) via FiLM/adapters and cross-attention. Reduced domain shift across devices/sites and improved generalization on held-out cohorts without extra labels.
• Reconstructed brain stiffness maps from 3D shear-wave data with GANs, ViTs, and diffusion models as learned inversions (vs. classical Helmholtz/phase-gradient methods). Achieved lower error and higher structural fidelity (MAE/SSIM) while cutting inference latency to enable near–real-time visualization and integrated per-voxel uncertainty for safer clinical interpretation.
• Designed ViT, diffusion, and UNet-family models for breast cancer segmentation under resource constraints: semi-/weakly-supervised learning with consistency training and pseudo-labels, class-imbalance handling, and deployment optimizations (ONNX/OpenVINO, INT8 quantization, pruning). Targeted low-memory edge workstations typical of screening/triage workflows.
• Improved physics and signal processing based, H-Scan ultrasound analysis for abnormality detection by refining backscatter spectral-shift estimation, attenuation compensation, and band-limited matched filtering. Stabilized maps across depth/gain settings and combined physics features with DL heads for robust lesion highlighting.
• Built an automated ultrasound registration/fusion + ML pipeline for cardiac and soft-tissue monitoring, cutting scan-to-diagnosis time by 37% and boosting detection F1 by +9 points (0.80→0.89) across ~160 patients (~1.3K studies). Reduced inter-operator variance by 31% and achieved AUC=0.93 on held-out validation.
• Developed GPU-accelerated Harmonic Motion Imaging (HMI) for tumor imaging and treatment monitoring, detecting ≥15% stiffness change within 10 days of therapy start. Improved SNR by +8.7 dB, achieved 5.4× faster runtime, and reduced non-diagnostic scans by 18% through real-time (20 fps) motion-artifact suppression.
• Key Skills: Deep Learning (PyTorch), Multimodal DL, Generative Models (GAN/ViT/Diffusion), Model Optimization & Quantization (ONNX/OpenVINO/INT8), Signal Processing, Quantitative Imaging, OCT/OCE, Elastography, Ultrasound Data Acquisition, DICOM/Clinical Integration, Experimental Design & Validation.

MSc Research Assistant — Shahid Beheshti University | 09/2018–01/2021

• Master’s Thesis — FPGA Super-Resolution (SR): Designed a fixed-point SR pipeline (deblurring → upsampling → refinement) on FPGA, using line buffers/BRAM tiling, DSP slice optimization, and AXI4-Stream DMA. Achieved stable timing closure at video rates; integrated camera input (MIPI/parallel) and LCD/HDMI output for on-device demos.
• Real-Time Stereo Vision on FPGA platform: Built a calibrated dual-camera pipeline (stereo K, D, R, T; rectification) and computed disparity using a graph-cut energy model (E = data + λ·smoothness) with α-expansion/β-swap moves, robust Potts smoothness, and occlusion handling via left–right consistency. Added sub-pixel refinement, median/bilateral filtering, and speckle removal. Converted disparity to metric depth (Z = f·B/d) and reprojected (Q matrix) to a colorized point cloud; downstream steps included plane/RANSAC fitting, normal estimation, and mesh reconstruction (Poisson/Ball-Pivoting) with optional multi-frame alignment via ICP for obstacle/scene understanding. Achieved real-time performance with high accuracy through parallelized cost-volume construction and memory-aware scheduling—targeted for autonomous vehicle perception.
• FPGA-accelerated iris recognition (embedded vision): Implemented segmentation (integro-differential/Daugman), normalization, Gabor filtering, and template matching with Hamming distance in fixed-point. Achieved 60 FPS at 1.8 W with a 0.1% error rate via pipelined architectures, resource sharing, and on-chip memory reuse.
• Built end-to-end lung CT pipelines in MATLAB: DICOM ingestion, resampling, lung/airway preprocessing, and classical ML (SVM/Random Forest with radiomics) alongside CNN baselines (2.5D/3D UNet variants). Implemented cross-validation, class-imbalance handling, and calibration to produce clinician-readable candidate lists.
• Developed TensorFlow transfer-learning models (ResNet/DenseNet backbones) for breast cancer detection on chest X-rays with explainability (Integrated Gradients) and uncertainty estimates. Standardized CXR preprocessing (de-identification, normalization, bone suppression) and reported FROC/sensitivity@FP consistent with clinical evaluation practices.
• Collaborated with clinicians/engineers to define labeling protocols, built reproducible training/eval scripts, and packaged models for edge deployment (C/C++/OpenCV) with automated tests and documentation.
• Selected Projects (M.Sc.):
  • Image registration for SR (FPGA-friendly): Registration pipeline tailored to fixed-point and on-chip memory for multi-frame SR (Winter 2021).
  • Implementation of LOG filter with new hardware friendly architecture for image enhancement on FPGA platform, Winter 2021 (Winter 2021).
  • FPGA-based image compression: DCT/quantization/entropy coding with AXI4-Stream I/O (Fall 2020).
  • Canny edge on FPGA (brain tumor borders): Fixed-point gradients/thresholding with line-buffer streaming (Spring 2020).
  • Super-Resolution in VHDL: Deblur–upsample–refine SR blocks with BRAM tiling and AXI-Stream I/O (Spring 2020).
  • Mano computer in VHDL: Complete educational CPU (datapath + control) simulation/implementation (Fall 2019).
  • Real-time stereo vision (VHDL prototype): Early disparity→depth pipeline for AV use (Summer 2019).
  • Digital Frequency Synthesizer (DDS): VHDL design with ASIC flow (Spring 2019).
  • ASIP for linear interpolation: Domain-specific ISA & micro-architecture derived from Mano computer (Spring 2019).
  • UART & SPI controllers: RTL design and verification in VHDL (Fall 2018).
  • 16-bit adder (180nm, Cadence): Layout-level implementation with DRC/LVS sign-off (Fall 2018).
  • Autism-support robot study: Concept exploration and human-factors review (Fall 2018).
  • Spin-coater build for nanotech lab: Control, safety interlocks, and process uniformity (Summer 2017).
  • Wearable antenna (CST): Compact, body-area resilient antenna modeling (Fall 2016).
• Key Skills: Machine/Deep Learning (MATLAB, TensorFlow), 2D/2.5D/3D CNNs, Radiomics, Explainable AI, Medical Image Processing (CT/CXR, DICOM), FPGA Design (VHDL/Verilog, HLS), Vivado/Quartus, Fixed-Point Quantization, DSP/BRAM optimization, AXI4/DMAs, OpenCV Optimization, Real-time Embedded & Edge Computing.

Selected publications

• Khaledyan, Donya, et al. “WATUNet: a deep neural network for segmentation of volumetric sweep imaging ultrasound.” Machine Learning: Science and Technology 5(1), 015042 (2024). View
• Khaledyan, Donya, et al. “Enhancing breast ultrasound segmentation through fine-tuning and optimization techniques: Sharp Attention U-Net.” PLOS ONE 18(12): e0289195 (2023). View
• Khaledyan, Donya, et al. “Low-cost implementation of bilinear and bicubic image interpolation for real-time image super-resolution.” In: Proc. IEEE Global Humanitarian Technology Conference (GHTC), 2020. View

Education

• Ph.D. Electrical & Computer Engineering | University of Rochester | 2022 – 2025
• M.Sc. Electrical & Computer Engineering | University of Rochester | 2022 – 2023
• M.Sc. Electronics Digital Engineering | Shahid Beheshti University | 2018 – 2021
• B.Sc. Electrical & Computer Engineering | Shiraz University of Technology | 2014 – 2018

Teaching Experience

Teaching Assistant | University of Rochester & Shahid Beheshti University | 2020 – 2023

• Served as TA for several courses, strengthening technical communication skills and deepening subject mastery through teaching: Medical Imaging, Signal Processing, FPGA, VLSI, Image Processing, Physics Electronics
• Skills Developed: Conceptual reinforcement, adaptive instruction, curriculum development.
Top