▸ About & CV
Shubham Somnath Sahoo
Building diffusion models, LoRA adapters & distilled on-device generative systems at Snap. IIT Kharagpur dual degree in EE & CS.
▸ Experience
Dec 2025 – Present
London
Machine Learning Engineer
Snap Inc.
- —Fine-tuned FLUX, QWEN diffusion models for Snap Lenses using LoRA and PEFT adapters for domain-specific Image2Image tasks.
- —Built a modular, config-driven prompt mining framework for image edit, style, video, and txt2image generation with taxonomy generation from specs.
- —Compressed teacher diffusion models into student GANs via kernel alignment and adversarial training for real-time mobile lens effects.
- —Distilled diffusion models into NAS-searched GAN architectures via kernel alignment, preserving latent structure for on-device inference, piloting face effect Lenses across 1M+ users.
FLUXQWENLoRAPEFTGANNASDistillationMobile ML
Dec 2023 – Aug 2025
Deep Learning Research Engineer
NeuroPixel.AI Labs
- —Implemented flat-lay to model image translation with FitDiT; refined apparel masks for garment alignment and realism.
- —Optimised Meta's Segment Anything (SAM) to TensorRT format achieving 2.69× faster inference for real-time segmentation.
- —Built a preemptive Redis queue to prioritise batch and real-time image generation requests by time, size, and urgency.
FitDiTSAMTensorRTRedisFastAPI
May 2022 – Nov 2023
Software Engineer
Analog Devices India
- —Reduced in-vehicle speech miss rate to 0.121% (at 72 km/h) with MIPS value of 13 for Voice Activity Detection.
- —Integrated Dolby Atmos In-Car Experience decoder using DMA interrupts and GPIO in a synchronous network.
- —Led deployment of non-contiguous memory allocation techniques, managing memory fragmentation and resource constraints.
DSPC++DMAMIPSDolby Atmos
Jun 2020 – Aug 2020
Research Intern
RPAD Lab, Carnegie Mellon University
- —Worked on Safety Envelope Tracking using self-supervised reward function and Graph Convolutional Neural Networks.
- —Devised approach to use temporal knowledge from environment by clustering points and greedy matching between frames.
- —Obtained over 80% accuracy with CNNs on residual policy prediction; optimised velocity baseline with ICP matching.
Self-supervised RLGCNICPCNN
May 2019 – Jul 2019
Robotics Intern
AMX Innovation, Bengaluru
- —Improved aerial detection model using transfer learning on YOLO V3-tiny object detection.
- —Integrated object avoidance planner for UAV manoeuvring and integrated code with ROS.
YOLO v3Transfer LearningROSUAV
▸ Skills & Expertise
Generative AI
FLUXSDXLQWEN Image EditLoRAPEFTAdaptersDiffusersStable Diffusion
Model Compression
Knowledge DistillationGlobal Kernel AlignmentNASGANQuantisationPruning
Computer Vision
Image2ImageSegmentationSAMFitDiTYOLOObject Detection
Deployment
ONNXTensorRTCUDAMobile MLKubernetesGCPAzure
Frameworks
PyTorchOpenCVFastAPILangGraphKafkaRedis
Languages
PythonC++CUDATypeScript
▸ Education
Jul 2017 – May 2022
IIT Kharagpur
Dual Degree · EE&C Engg + CS Minor
Electronics and Electrical Communication Engineering · M.Tech in Visual Information Processing (M.Tech) · Minor in Computer Science and Engineering
CGPA: 8.65 / 10.0
▸ Publications
COMSNETS 2022
Efficient last-mile link adaptation in next-gen wireless heterogeneous networks
DOI: 10.1109/COMSNETS53615.2022.9668503
PhD EDITS 2020
A novel machine learning approach for link adaptation in 5G wireless networks
DOI: 10.1109/PhDEDITS51180.2020.9315299
ICCRT 2019
A prototype of an intelligent ground vehicle for constrained environment: Design and development
DOI: 10.1145/3387304.3387321