class: research_scientist · conf: 0.93 | dreamer · conf: 0.99

Ashuta Bhattarai

Research Scientist & ML Engineer

Deep learning · Computer vision · Generative modeling · Applied AI systems

Ashuta Bhattarai
[ 01 · Introduction ]

Hi, I'm Ashuta. I'm a research scientist and ML engineer who spends most days deep in computer vision, generative modeling, and applied AI, but I like to think there's more to me than the models I train. I did my PhD at the University of Delaware, and along the way I've worked on everything from medical imaging to video synthesis to red-teaming large language models. Outside of work, I'm usually curled up with a good novel, chasing down a tricky puzzle, or arguing over triple word scores in a game of Scrabble. I love to travel whenever I get the chance, and I have a quieter side that leans toward literature, poetry, and art. I sometimes let that side show on Medium, where I write a bit differently than I do in a research paper.

[ 02 · My work life so far... ]

Research Scientist Consultant

May 2026 – Present
Fusemachines Inc.
  • Led research for a real-time interactive video avatar agent for AI-driven interview, involving a fully expressive, lip-synced visual experience using diffusion-based generative models for photorealistic rendering.
  • Collaborated with a cross-functional research team to define feasibility, model selection, and evaluation benchmarks for production deployment of the live video agent, reducing interactive conversation latency by 50% in a live meeting.
  • Upgraded a text-based LLM pipeline of a live interview agent with a frontier audio-native model, cutting latency by ~39% and improving LLM-as-judge accuracy by ~3%.

Machine Learning Engineer Consultant

Aug 2025 – May 2026
Quasar Intelligence
  • Architected the end-to-end Bone Health Index prediction pipeline spanning phantomless HU calibration, 3D preprocessing, multi-domain segmentation, and risk score generation from volumetric CT scans across 4 major scanner vendors, with sub-60-second inference.
  • Developed a longitudinal clinical biomarker quantification system using composite features such as bone mineral density, skeletal muscle mass, and IMAT from CT scans across diverse patient populations, enabling structured tracking of musculoskeletal health indicators over time.
  • Built a scalable, reproducible ML workflow using PyTorch, and MONAI, processing hundreds of CT scans with standardized cross-vendor compatibility, enabling downstream clinical validation and research deployment.
  • Designed an LLM-powered clinical reporting pipeline that ingests model-generated measurements (bone density, muscle mass, visceral fat) and produces structured, per-patient BHI narrative reports via prompt engineering.

Graduate Research Assistant

Feb 2020 – May 2025
Video/Image Modeling and Synthesis (VIMS) Lab · University of Delaware, Newark, DE
  • Developed a self-supervised pretraining algorithm for video understanding by reconstructing temporally shuffled frames, matching VideoMAE performance on action recognition benchmarks without requiring intact temporal sequences.
  • Built a deep learning pipeline for Sickle-Cell Retinopathy detection from OCT scans with Nemours Children's Hospital, achieving mAP 0.86 on 500+ clinical volumes (published at MICCAI 2024 as top 11% of 2,869 submissions).
  • Investigated transformer and CNN-based architectures for retinal layer segmentation and contour detection in clinical OCT imaging, contributing to reproducible and quantitative biomarker extraction from ophthalmic data.
  • Mentored junior researchers on deep learning methodology, experimental design, and PyTorch workflows, enabling independent experimentation and contributing to 2 additional conference submissions.

Machine Learning Engineer

Jan 2018 – Jan 2020
Ekbana Solutions · Lalitpur, Nepal
  • Developed instance segmentation and object detection pipelines for satellite imagery using Mask RCNN and RetinaNet, building end to end workflows from data preparation to model inference for ship detection in large scale aerial scenes.
  • Built time series forecasting models for retail sales using PyTorch and Scikit learn, predicting demand patterns that guided store inventory and stocking decisions, improving supply chain efficiency.
[ 03 · Education ]

PhD, Computer Science (M.S. en route)

University of Delaware · Newark, DE
Feb 2020 – May 2025

B.E., Computer Engineering

Tribhuwan University · Lalitpur, Nepal
Nov 2013 – Sep 2017
[ 04 · Publications ]
  1. [1] Bhattarai A, Jin J, and Kambhamettu C. Analyzing Adjacent B-Scans to Localize Sickle Cell Retinopathy in OCTs. Medical Image Computing and Computer Assisted Intervention (MICCAI), 2024.
  2. [2] Bhattarai A, Kambhamettu C, and Jin J. CUNet: Towards continuous multi-class contour detection for retinal layer segmentation in OCT images. 29th IEEE International Conference on Image Processing (ICIP), 2022.
  3. [3] Bhattarai A. Towards Multi-Scale Inter-Frame Attention to Improve Deep Learning Tasks. Ph.D. dissertation, Dept. of Computer and Information Sciences, University of Delaware, Newark, DE, USA, 2025.
[ 05 · Technical Skills ]
Languages
PythonMATLABC++SQL
Machine Learning
Deep LearningComputer VisionMedical Image AnalysisVideo UnderstandingObject DetectionInstance SegmentationSelf-Supervised LearningLLM Red-TeamingPrompt EngineeringTime Series ForecastingDiffusion Models
Frameworks & Libraries
PyTorchTensorFlowMONAIHuggingFace TransformersNumPyPandasOpenCVScikit-learn
MLOps & Tools
Weights & BiasesTensorBoardGitLinuxAWS SageMaker