Pooja Ravi

Machine Learning Engineer
MScAC (AI) @ University of Toronto · Vector Scholar in AI
Toronto, Canada · kuchan2001@gmail.com
Google Scholar GitHub Kaggle Medium LinkedIn

About

I am a machine learning engineer with a focus on vision-based systems, and model performance optimization. My work spans researching, training and deploying generative models, developing multimodal pipelines, and model performance optimization. I was awarded the Vector AI Scholarship for 2023-24, published four peer-reviewed papers during undergrad and hold a patent-pending algorithm for visual representation learning.

Research Interests: Generative Modeling, 3D Vision, Model Optimization, Representation Learning

Education

Master of Science in Applied Computing (MScAC)
University of Toronto
Sep 2023 – Jan 2025
Vector Scholarship in AI · Mitacs Globalink Graduate Fellowship
Courses: Computational Imaging (D. Lindell), ML for Healthcare (R. G. Krishnan), Natural Language Computing (G. Penn), Visual & Mobile Computing (N. Vijaykumar)
B.Tech. Computer Science & Engineering
SRM Institute of Science and Technology, Chennai, India
Jun 2019 – May 2023

Publications

[1] A. Shukla, P. Ravi, and B. Arthi. "An integrated document manipulation tool leveraging LLMs, transformers, and attention mechanism." AIP Conference Proceedings, April 2023. [DOI]
[2] P. Ravi and A. Dewan. "Real-time Multi-Module Student Engagement Detection System." ICCIS, NIT Delhi, December 2022. LNNS, vol. 686, pp. 261-278. [DOI]
[3] P. Ravi, S. Roy, I. Dutta, and K. Kottursamy. "Attention Mechanism, Linked Networks, and Pyramid Pooling Enabled 3D Biomedical Image Segmentation." IEEE/ACIS 23rd International Conference SNPD, Taiwan, 2022, pp. 91-96. [DOI]
[4] P. Ravi, A. Shukla, and B. Muruganantham. "Real-time GPU-accelerated Driver Assistance System." RCAAI, November 2022. LNEE, vol. 1066, pp. 821-834. [DOI]

Noteworthy Projects

EchoCharlieBoson AI Hackathon Winner, Oct 2025
A multimodal voice restoration pipeline that rebuilds lost or corrupted audio in videos using visual speech recognition, transcript cleanup, and AI-driven voice cloning models.
Multimodal Voice Cloning Speech Recognition Video Processing
NeRF: 2D Sketch to 3D Stylization
Convert 2D sketches to stylized 3D views using neural radiance fields. Image style transfer with NeRF trained from scratch (few-shot learning) and CLIP-NeRF for text prompt-based stylization.
NeRF CLIP Style Transfer Few-Shot
MobileGen: Optimizing Diffusion Models
Optimized DDPM for mobile inference via quantization-aware training, knowledge distillation, pruning, and U-Net architecture modifications. 20% faster inference, 14% lower GFLOPs.
DDPM Quantization Distillation Mobile
RayGen: X-ray Image Synthesis
Cross-modal pipeline aligning vision and language embeddings to reconstruct latent representations for chest X-ray synthesis. Masked images concatenated with BERT-processed radiology reports.
Cross-Modal BERT Medical AI Synthesis
Sketch2Shoe with Pix2Pix GAN
Pix2Pix GAN implemented from scratch with mixed-precision training and post-training pruning. Converts grayscale shoe sketches to photo-realistic versions.
Pix2Pix GAN Mixed Precision Pruning
EasyDocs
All-in-one academic utility: text summarization, document OCR/parsing, keyword extraction, similar-publications recommender, and text-to-speech.
NLP OCR TTS Recommender

Experience

Machine Learning Engineer May 2024 – Present
Q2 Software Inc., Austin, USA (Remote)
  • Independently led the signature verification project: scalable end-to-end workflows for cheque signature comparison across a 200k image database. Trained variational autoencoders for signature embeddings, used triplet loss with contrastive learning and Siamese networks, and fine-tuned multimodal LLMs for real-time visual similarity assessment.
  • Contributed to an end-to-end cheque fraud prevention pipeline. Trained LayoutLM, Swin+BART (Donut), and TrOCR on 100k+ cheque images. Improved text extraction by 8%, flagging rate by 4%, and reduced inference latency. System in production for 270+ financial institutions.
  • Patent Pending — Scalable visual representation learning algorithm with pretrained feature extractors and high-dimensional clustering for cheque format diversity.
Machine Learning Engineer Oct 2023 – Mar 2024
16 Bit, Toronto, Canada (Part-time)
  • Led research on predicting knee osteoarthritis progression (~96 months) using image-to-image translation on X-ray images depicting different arthritis grades.
  • Worked with GANs, contrastive learning, diffusion models, PyTorch Lightning distributed training, and LoRA. Achieved ~10x training speedup for 70k+ X-rays and ~19% increase in image realism.
  • Translated generated images into clinically relevant output mappings with medical domain experts.
Computer Vision Intern Jun 2021 – Sep 2021
Blunav Technologies, IIT Madras, India
  • Automated aircraft turnaround detection from real-time CCTV feeds using AWS SageMaker, object detection, tracking, and in-video distance calculation.
  • 3x real-time inference speedup and 11% mAP improvement with scaled-YOLOv4.

Research

Mitacs Globalink Research Intern Jun 2022 – Sep 2022
Athabasca University, Edmonton, Canada — Supervised by Dr. Ali Dewan
  • Built a real-time student engagement detection pipeline using 3D residual, spatiotemporal convolutional networks for video analysis.
  • Frame classification and 3D mesh facial landmark tracking: ~10.4% increase in tracking performance, up to 12% latency decrease via multiprocessing.
Undergraduate Researcher Mar 2021 – Oct 2022
SRM IST, Chennai, India — Supervised by Dr. B. Muruganantham & Dr. Kottilingam K
  • Driver safety monitoring system using Swin Transformer for road entity detection, physical distance estimation, and microsleep/drowsiness tracking.
  • 3D image segmentation networks with soft attention, residual blocks, and pyramid pooling, benchmarked on BraTS2020 for brain tumor localization.

Technical Skills

Core: Python, PyTorch, AWS SageMaker, Docker, JAX, SQL
Libraries: OpenCV, NumPy, HuggingFace, TensorFlow, NLTK, TensorRT
Vision: Object Tracking, Semantic Segmentation, Video Diffusion, OCR 3D Vision
Optimization: Quantization, Pruning, Distributed Training, LoRA, Knowledge Distillation
Healthcare: Medical Imaging, X-Ray Synthesis, 3D Segmentation, Time Series