Open to Research & Industry Roles

Mayesh Mohapatra

Working in

Graduate Researcher at Texas A&M working on generative models for image synthesis and 3D vision. Interested in diffusion transformers, flow matching, and controllable generation. Published at IEEE IGARSS 2025 and ISRO-IISc. Previously shipped ML systems at IISc and Deloitte that cut inference latency by 10,000×.

Download Resume View Projects

Experience

Graduate Researcher — 3D Perception
Texas A&M University  ·  Advisor: Prof. Cheng Zhang  ·  College Station, TX
Aug 2025 – Present
  • Built an end-to-end monocular 3D scene reconstruction pipeline for dynamic athlete reconstruction from single-view RGB video, integrating Human3R and VGGT for pose estimation with DeepLabv3 segmentation into a unified PyTorch tracking framework; validated on internal multi-athlete benchmarks.
  • Cut end-to-end inference latency by 40% for real-time biomechanical feedback during live training sessions, using mixed-precision CUDA kernels and INT8 quantization across the full perception stack.
  • Built an analytics module that extracts biomechanical metrics (jump height, center-of-mass velocity, take-off angle) across 3+ athletic disciplines by fusing 3D pose outputs with physics-based motion analysis.
  • Delivered production perception APIs consumed by biomechanics researchers and coaching staff, with standardized data contracts linking 3D reconstruction outputs to a downstream decision-support system.
Project Associate — Multi-Modal Perception & Satellite Image Analysis
Indian Institute of Science (IISc) & ISRO Space Applications Centre  ·  Bengaluru, India
May 2023 – May 2025
  • Boosted model training throughput by 30% on 100+ GB/day of HDF5 multispectral satellite imagery by building an automated Python ingestion pipeline that generated roughly 14,000 annotated frames via GDAL and Rasterio.
  • Designed a two-stage cyclone detector (Center-Locator and Intensity-Estimator) in PyTorch with self-supervised pre-training on unlabeled satellite data, validated across 5+ spectral sensor bands using mAP and RMSE.
  • Reduced inference latency from 3–4 hours to under 1 second (10,000× speedup) for near-real-time cyclone monitoring, applying model pruning, INT8 quantization, and parallelized preprocessing.
  • Scaled training to 200+ tracked experiments with reproducible deployments using Weights & Biases, Docker, and distributed Spark/Dask data workflows.
  • Published at IEEE IGARSS 2025 (Brisbane) on deep-learning models for cyclone detection on infrared satellite imagery.
Research Intern (Remote) — Generative Models & Anomaly Detection
Georgia Institute of Technology  ·  Atlanta, GA
Jan 2023 – May 2023
  • Improved out-of-distribution robustness by 15% on insider-threat detection by synthesizing rare-class training data with GANs in PyTorch.
  • Achieved 95% classification accuracy and lifted rare-event recall by 12%, training autoencoder-based anomaly detectors combined with a Random Forest and Logistic Regression ensemble.
  • Cut data preprocessing time by 25% across 5+ dataset splits by vectorizing cleaning workflows in Pandas and NumPy.
Data Science Intern — ML Pipelines & Analytics
Deloitte Haskins & Sells LLP  ·  Bengaluru, India
Jun 2022 – Mar 2023
  • Saved 15 engineering hours/week and cut audit review time by 40% by refactoring Spark ETL pipelines and overhauling journal-entry testing logic across distributed data clusters.
  • Deployed analytics dashboards adopted by 4+ audit teams (85% deployment rate) by building Power BI exception-analysis tools and automating reporting workflows.

Education

Texas A&M University
M.S. Computer Science  ·  College Station, TX
GPA: 3.67 / 4.0 Aug 2025 – May 2027
Deep Learning CV & Robotic Perception ML for 3D Vision & Graphics Deep Reinforcement Learning Generative AI for CV Advanced CV
Bennett University
B.Tech (Honors), Computer Science Engineering  ·  Greater Noida, India
GPA: 3.54 / 4.0 May 2023
Machine Learning Data Structures & Algorithms Linear Algebra Probability & Statistics Database Systems

Projects

Feb 2026 – Present
Manuscript in Progress
Hierarchically Decomposed Diffusion Guidance (HDG)
  • Co-developed HDG, a method decomposing classifier-free guidance into hierarchical components for fine-grained conditional diffusion generation, improving stability, sample quality, and controllability over standard CFG.
  • Built data pipelines aligning fine-grained class hierarchies with diffusion conditioning signals across multiple semantic taxonomies.
  • Implemented full evaluation suite measuring generation quality, controllability, and fine-grained alignment with ablation studies isolating each decomposition component vs standard CFG baselines.
Diffusion Models CFG PyTorch Flow Matching
Jan – May 2026
Completed
Pose4DGS: Pose-Conditioned 4D Gaussian Splatting
  • Improved unseen-pose generalization on monocular 4D Gaussian Splatting human avatars with two zero-overhead training modules: SMPL-based Pose-Space Augmentation and DINOv2 (ViT-B/14) feature supervision.
  • Designed a 3-tier novel-pose evaluation protocol (easy / moderate / extreme) on ZJU-MoCap and PeopleSnapshot; ran full 2×2 ablation grid; compressed DINOv2 embeddings 768→64 via PCA.
  • Led a 3-person team across rendering, deformation, and evaluation modules; tracked all ablations in W&B; delivered a reproducible 6–8 page technical report.
4D Gaussian Splatting DINOv2 SMPL CUDA W&B
Aug 2025 – Present
Active Research
Monocular 3D Human Reconstruction
  • Building a neural rendering pipeline converting single-view RGB video into pose-driven 3D human meshes using neural implicit representations and SMPL priors, targeting sub-centimeter surface accuracy.
  • Benchmarked reconstruction quality against SOTA baselines on PSNR, SSIM, and LPIPS on ZJU-MoCap and Human3.6M by iterating on deformation-network architectures and loss functions.
NeRF Gaussian Splatting SMPL COLMAP PyTorch
May 2023 – May 2025
IEEE IGARSS 2025
CycloAI: Multi-Modal Cyclone Detection
  • Co-built end-to-end perception pipeline processing 100+ GB/day satellite streams with a two-stage detector (Center-Locator + Intensity-Estimator) with self-supervised pre-training; reduced inference from 3–4 hrs to <1 sec.
  • Lifted minority-class recall by 12% across 5+ spectral bands using focal loss and SMOTE; tracked 200+ runs in W&B inside Docker-containerized environments.
PyTorch GDAL Spark/Dask Docker CUDA
Aug – Dec 2025
Completed
Recurrent Attention Model (RAM)
  • Achieved 96.64% MNIST accuracy using only 6 glimpses (~49% of pixels) by reproducing RAM in PyTorch with REINFORCE policy gradients; matching full-image CNN baselines at a fraction of the compute.
  • Stabilized policy-gradient training with a learned value baseline; ablated glimpse count, patch resolution, and reward shaping; released a reproducible open-source codebase.
REINFORCE PyTorch Attention NumPy
Aug – Dec 2022
Completed
CNN Architectures from Scratch
  • Implemented ResNet and VGG variants from first principles in PyTorch without pre-trained weights; assessed convergence on CIFAR-10 and ImageNet subsets with gradient-flow analysis.
  • Diagnosed vanishing gradient issues in deep VGG, demonstrating ~2× faster convergence with skip connections in ResNet counterparts.
ResNet VGG PyTorch NumPy

Publications

01
Comparative Analysis of Deep Learning Models for Cyclone Detection and Localisation on Infrared Satellite Images
IEEE IGARSS 2025, Brisbane, Australia IEEE
02
A Self-Supervised Visual Analytical Approach for Automatic Detection of Cyclonic Events in Satellite Observations
40th Annual Symposium on Space Science and Technology, ISRO-IISc, 2024 arXiv
03
GAN-Based Approach for Synthetically Generating Insider Threat Scenarios
Journal of Software Engineering and Applications, 16(11):586–604, 2023 Journal

Skills

Programming

Python CUDA C++ SQL R

ML Frameworks

PyTorch HuggingFace Diffusers / Transformers Scikit-Learn NumPy / Pandas Weights & Biases

Generative Models

Diffusion Transformers (DiT, FLUX.1, SD3) Flow Matching / Rectified Flow DDPM / DDIM Classifier-Free Guidance VAE Score-Based Models

Computer Vision

3D Scene Reconstruction Gaussian Splatting NeRF Multi-View Geometry Human Pose Estimation SMPL / SMPLx Semantic Segmentation Point Cloud Processing

Vision Backbones

DINOv2 (ViT-B/14) Vision Transformers CLIP

Systems & Tools

Docker COLMAP GDAL / Rasterio Spark / Dask Git Linux JupyterLab

Let's Connect

Based in College Station, TX. Actively seeking research internships and full-time roles in Computer Vision, 3D Reconstruction, Neural Rendering, and Generative Models. Open to both academic labs and industry research teams.

mayesh@tamu.edu LinkedIn Full Resume
© 2026 Mayesh Mohapatra College Station, TX