Surya Santosh Kumar Yellapu

• Designed and implemented quantization modules for TI Deep Learning (TIDL) framework, enabling low-precision (INT8/INT16) inference with minimal accuracy loss.
• Specialized in Post-Training Quantization (PTQ), Quantize-Dequantize (QDQ), and Mixed-Precision inference for complex models including Transformers, object detection, and segmentation networks.
• Implemented support for proto-based model representations (prototxt, protobuf) to enable scalable mixed-precision execution and seamless model introspection via Python interfaces.
• Built containerized TIDL development environments to support diverse developer workflows and CI/CD integration.
• Have experience working directly with enterprise clients (e.g. Ford, Valeo) on quantization and accuracy issues.
• Mentored junior developers and interns in building modules for streamlined pipeline analysis and performance optimization of operators and neural networks.

• Built a diffusion-based object removal pipeline based on Stable Diffusion SDv1.5 Inpainting model.
• Designed a custom mask-guided decoder for background coherence and accelerated inference to under 0.2s using LCM distillation with LoRA optimization.
• Trained the MobileFaceNet and Resnet50 models based on Arcface to scale over millions of faces.
• Achieved a 97% model size reduction (to 1 MB) by combining structured pruning, NAS, and quantization making it possible to deploy on mobile hardware.
• Compressed UNet and DiT diffusion models using SVD for attention matrices, reducing model size by 50% and inference time by 20% with only a 0.001 increase in FID.
• Designed a novel training pipeline for fine object selection using only bounding boxes, outperforming SOTA models.
• Distilled and quantized student ViTMatte model for edge deployment with 66% lower memory bandwidth and 33% faster inference under 100 ms.

• Designed and deployed a WebRTC media application using Kurento Media Server on a Node.js backend with MongoDB and containerised the application using docker.
• Built a real-time internal analytics dashboard using React and Firebase for monitoring system metrics.
• Created a Dockerized microservice using Flask, OpenCV, and FFmpeg to offer cloud-based image processing APIs such as compression, format conversion, image filters.

Indian Institute of Technology, Kharagpur (IIT KGP)

Bachelors of Technology (Honours) in Electronics and Electrical Communication Engineering

Prajwal G, Surya Santosh Kumar Y, NKS Reddy, Prabhath C, Avinash T, Neeraj K, C Shyam Anand

Published in the Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2023.

Developed a real-time safety system to prevent human-machine collisions using object detection, trajectory prediction, and monocular depth estimation.

Created a python library for vector operations and vector calculus in 2D and 3D space.

Built a python library XWindows to access the Xserver on Linux for accessing raw video stream information and for sending the responses back.