Gaurav Goswami
ML Engineer @ Red Hat | VLMs & LLMs | Quantization (GPTQ/AWQ) | vLLM | PyTorch | Edge AI
Bengaluru
8+ years of work experience
Sep 20251 yr 1 moPresent
Bengaluru
• Building and optimizing vision-language model (VLM) architectures with efficient attention mechanisms for multimodal AI workloads
• Driving quantization research (GPTQ, AWQ, SmoothQuant, QAT) to compress large transformer and diffusion models for low-latency deployment
• Contributing to the PyTorch ecosystem and open-source AI tooling for scalable model development
• Working with vLLM to optimize high-throughput inference for large language and vision-language models
• Developing scalable training workflows using Training Hub for large-scale model training pipelines
Mar 20241 yr 7 mosSep 2025
Bengaluru
• Developed and deployed high-resolution vision models for Galaxy S24, S23, and Z Fold series — shipped on 100M+ consumer devices across 3 flagship generations
• Led end-to-end AI pipeline for real-time on-device inference, optimizing vision transformer and CNN architectures for Samsung Exynos NPUs and Qualcomm DSPs
• Drove quantization-aware training (QAT) and post-training quantization (PTQ) with per-channel calibration and mixed-precision (INT8/FP16) scheduling to reduce latency 3–5× without accuracy loss
• Explored on-device vision-language model (VLM) architectures for adaptive scene understanding and intelligent camera control — prototyped lightweight CLIP-style encoders fused with task-specific heads
• Built Bayer-domain video denoising and NV21 low-light enhancement modules, improving dynamic range and visibility in extreme illumination — integrated into camera HAL layer
• Evaluated diffusion-based restoration methods and designed latent-space quantization strategies to make generative models viable for edge deployment
Oct 20212 yrs 6 mosMar 2024
Bangalore Urban
• Developed low-light image enhancement and denoising models with hardware-aware optimization, deployed across mobile and embedded platforms with INT8 quantization for real-time inference
• Built AI-based denoising, enhancement, and Bayer demosaicing pipelines optimized for NPU/DSP execution — profiled memory bandwidth and operator rooflines for bottleneck analysis
• Implemented model compression via structured pruning, knowledge distillation, and weight sharing — delivering <15ms per-frame latency on mobile SoCs
• Built and optimized object detection and segmentation pipelines with quantization-friendly architectures tailored for efficient on-device inference
• Developed reflection artifact removal and moire-removal algorithms integrated into production camera pipeline for improved image quality
• Contributed to Long Range Face Detection with efficient attention mechanisms, improving subject recognition at extended distances on Samsung flagship devices
Nov 20192 yrsOct 2021
• Vehicle Detection & Tracking — Designed custom feature-pyramid detector optimized for multi-class vehicle detection with bounding-box post-sorting and motion-assisted temporal filtering for robust multi-frame tracking; deployed on Jetson Nano with channel-level pruning and ONNX export for real-time edge inference
• ALPR System — Developed anchor-based detector for license plate localization with multi-head OCR classifier and CTC-based decoding; applied post-training quantization (INT8) for deployment on Jetson TX2 at 30 FPS; integrated rule-based logic for helmet detection, wrong-way tracking, and triple-rider analysis
• Document OCR Pipeline — Designed lightweight CNN + RNN pipeline for structured text extraction from PAN, Aadhaar, and license images with alignment correction and ROI extraction; optimized with INT8 quantization and deployed on Jetson Nano for real-time document verification
Jul 20187 yrs 2 mosAug 2025
Remote
• Reparameterized Denoising Network — Built multi-branch residual attention architecture with dynamic convolutional paths and channel re-calibration on SIDD; designed reparameterization-friendly blocks compatible with post-training quantization (PTQ) for single-pass inference fusion at deployment
• PyTorch Quantization Framework — Engineered end-to-end QAT framework supporting per-channel symmetric/asymmetric quantization, fake quantization ops, observer modules (MinMax, Histogram), and post-training static quantization; benchmarked accuracy-latency trade-offs across INT8/INT4 precision levels
• YOLOE Object Detection — Implemented YOLOE with enhanced E-ELAN backbone, decoupled detection head, and anchor-based target assignment; designed INT8-quantization-friendly architecture with COCO-style evaluation and TensorRT export path
• YOLOX PyTorch — Built anchor-free detector with CSPDarkNet backbone, PA-FPN, decoupled head, and SimOTA label assignment; optimized for real-time edge inference with ONNX export, operator fusion, and mixed-precision (FP16/INT8) deployment support
• Low-Level Vision Suite — Curated modular implementations of SRCNN, DnCNN, and DeblurGAN with perceptual loss and pixel shuffle upsampling; added quantization-aware training paths and CLIP-based perceptual quality metrics alongside PSNR/SSIM evaluation
• TeethU2Net — Customized U2-Net for dental image segmentation with multi-scale side outputs, residual U-blocks, and binary focal loss; evaluated FP16/INT8 ONNX export paths for clinical edge device deployment
• Object Localization Network — Developed regression-based localization CNN with smooth L1 loss, spatial dropout, and adaptive feature extraction; benchmarked on Pascal VOC with IoU-based metrics and lightweight architecture design for mobile inference
Jul 20181 yr 5 mosNov 2019
Gurugram
• Traffic Incident Detection — Implemented rule-augmented pipeline to detect stopped vehicles, reverse motion, over-speeding, and environmental factors (fog, smog); deployed as streaming video processor using GStreamer + TorchScript on industrial edge setups
• Vehicle Axle Classification — Built custom HOG-feature + CNN hybrid models for axle detection from undercarriage views with morphological filtering for axle isolation and class distinction
• People Counting System — Developed background-subtraction-based counting system with bounding box merging and movement vector analysis; deployed in pure C++ using OpenCV on Raspberry Pi and RockChip Pro — optimized for <50MB memory footprint
• Synthetic Data with GANs — Trained Conditional GAN to generate number plate images with varied fonts, lighting, and occlusions; synthetic augmentation improved OCR accuracy by 12%+ on rare and degraded captures
Gaurav’s AI Stack
Education
Bachelor of Engineering, Computer Engineering
Mahatma Jyotiba Phule Rohilkhand University (MJPRU), Bareilly
20144 yrs2018
Join Gaurav and 10k+ other professionals on
Tal is the app where bosses meet engineers & builders directly, without middlemen
Download Tal Boss