Yaxita Amin

M.S. Applied Machine Learning · University of Maryland, College Park

Graduating 2026 · Open to Full-Time AI/ML Roles

I am a graduate researcher at the University of Maryland, College Park, working at the intersection of efficient deep learning, human preference alignment, and intelligent systems.

My research spans efficient attention mechanisms, multi-agent systems, and computer vision.

Previously, I was a Computer Vision Research Intern at VGEC, Gujarat Technological University, where I published first-author work on U-Net architectures for satellite change detection at IET PICET 2025.

yaxita2003@gmail.com linkedin.com/in/YaxitaAmin github.com/YaxitaAmin
College Park, MD, USA
CV GitHub LinkedIn arXiv
News
Publications
🤖
CHOP: Counterfactual Human Preference Labels Improve Obstacle Avoidance in Visuomotor Navigation Policies
Gershom Seneviratne, Jianyu An, Vaibhav Shende, Sahire Ellahy, Yaxita Amin, Kondapi Manasanjani, Samarth Chopra, Jonathan Deepak Kannan, Dinesh Manocha
arXiv:2603.02004 · University of Maryland, College Park · March 2026 · Targeting RSS 2026
arXiv PDF
🛰️
A Comparative Study of U-Net Architectures for Change Detection in Satellite Images
Yaxita Amin et al.
IET Conference Proceedings · PICET 2025 · Peer-Reviewed · First Author
Paper
Research Experience
Computer Vision Research Intern Jan 2024 – Jun 2024
Gujarat Technological University – VGEC
  • Comparative analysis of 18 U-Net architectures with attention and transformer modules for satellite change detection.
  • Designed modular ablation pipelines; developed accuracy–efficiency–complexity trade-off metrics.
  • First-author publication in IET Conference Proceedings (PICET 2025).
Selected Projects
MLP-Parameterized Lambda in Log-Linear Attention NeurIPS 2026 target
Extended Log-Linear Attention (ICLR 2026) by replacing the static scalar lambda with an input-dependent 2-layer MLP (hidden dims 32/64/128). Achieves 99.6%±0.1% vs baseline collapse to 60.9%±46.7% on MQAR, and 10× length generalization improvement. Phase 1 complete; language modeling on Wikitext-2 ongoing. Run on Zaratan HPC (H100 GPUs).
PyTorchTritonHPC · H100Sequence ModelingAttentionNeurIPS 2026
99.6% MQAR
10× length gen
GitHub ↗
Multi-Agent Android QA System
Planner → Executor → Verifier → Supervisor pipeline across OpenAI, Claude, and Gemini.
LangChainPythonStreamlit
87% success
Demo ↗
Surgical Robot RL Navigation
Q-Learning and DQN for autonomous needle navigation in 3D vasculature with Dockerized deployment.
PyTorchRLDocker
98–100% success
0% collision
BERT Model Compression Pipeline
Quantization, pruning, and knowledge distillation with automated CI/CD on AWS EC2.
PyTorchAWSCI/CD
80% size ↓
47% latency ↓
Autonomous Portfolio Risk Management
6-agent real-time monitoring with ML Anomaly Detector and Risk Scanner on live data streams.
PythonIsolation ForestStreamlit
Demo ↗
LeafDoctorAI — Agricultural Disease Detection
CNN-based plant disease classification with mobile-optimized inference on 10K+ images.
CNNPyTorchOpenCV
92% accuracy
ML Memory Profiling Framework
Comparative PyTorch vs TensorFlow memory analysis with interactive Streamlit dashboard.
PythonProfilingStreamlit
4.3× efficiency
Demo ↗
Skills
PythonPyTorchTensorFlowLangChainLLM Fine-tuningRLHF / SFTTransformersComputer VisionHugging FaceAWSDockerSpark MLlibOpenCVCI/CDScikit-learnStreamlitAblation StudiesPaper Writing