Bishram Acharya
Bishram Acharya

Bishram Acharya

Research Assistant · NAAMII — TOGAI Lab
Computer Engineering graduate (BE) from Pulchowk Campus, IOE, Tribhuvan University. Research Assistant at NAAMII's TOGAI Lab, developing novel AI methods for fetal ultrasound analysis in low-resource clinical settings. Research interests span medical imaging, vision-language models, and self-supervised learning.
Download CV Email LinkedIn GitHub

Education

Bachelor of Engineering in Computer Engineering
Pulchowk Campus, Institute of Engineering, Tribhuvan University · April 2021 – Oct 2025
  • Ranked 149th in the nationwide engineering entrance exam out of 15,000+ applicants (top 1.0%)
  • Graduated with First Division
  • Undergraduate Thesis: AI-Generated Image Detection via Frequency Domain Learning
  • Relevant Coursework: Artificial Intelligence, Probability & Statistics, Numerical Methods, Engineering Math I, II & III, Applied Math, Applied Data Science, Computer Graphics, Simulation & Modeling, Data Structures & Algorithms, Theory of Computation
Higher Secondary Education, Physics
St. Xavier's College, Maitighar · 2018 – 2020
  • Physics Stream with GPA 3.79/4.0. Strong foundation in Mathematics and Physics' first principles thinking

Publications

MIDL 2026 · Feb 2026
Automated Quality Assessment of Blind Sweep Obstetric Ultrasound for Improved Diagnosis
Prasiddha Bhandari, Nishant Luitel, Kanchan Poudel, Bishram Acharya, Angelina Ghimire, Tyler Wellman, Kilian Koepsell, Pradeep Raj Regmi, Bishesh Khanal  ·  Medical Imaging with Deep Learning
📄 Paper ↗
  • Systematically evaluated AI robustness to protocol deviations in sweep ultrasound videos; developed automated quality-assessment models to detect perturbations and improve downstream task performance
  • Simulated a clinical feedback loop demonstrating measurable gains across three classification tasks: sweep-tag, fetal presentation, and placenta location
  • Primary contribution: Ran scalable PyTorch Lightning and Hydra pipelines for systematic experimentation and reproducibility, while actively contributing to manuscript preparation

Professional Experience

Research Assistant (full-time)
NAAMII (TOGAI Lab) · Supervisor: Dr. Bishesh Khanal · June 2025 – Present
  • Core research direction: Developing novel AI methods to analyse fetal ultrasound videos for automated diagnosis in low-resource clinical settings
  • Contributing to industry R&D collaboration with GE Healthcare via NAAMII, working on solving real-world challenges of AI-assisted obstetric ultrasound
  • Monitoring and analysing a first-of-its-kind blind sweep fetal video dataset collection, conducted in collaboration with Institute of Medicine, Nepal
  • Research spans foundational models, domain adaptation across demographic datasets, label noise robustness, JEPA/world models, active learning, and multi-task learning
  • Co-authored MIDL 2026 paper on automated quality assessment of blind sweep obstetric ultrasound
Research Intern
NAAMII (TOGAI Lab) · Supervisor: Dr. Bishesh Khanal · Sept 2024 – May 2025
  • Built an end-to-end pipeline for the ISIC skin lesion classification challenge including ensemble methods and test-time augmentation
  • Co-developed a methodology combining Active Learning and Vision-Language Segmentation Models (VLSMs) for medical image annotation reduction; conducted systematic literature review and contributed to manuscript preparation
Teaching Assistant
Annual AI School (ANAIS), NAAMII · Dec 2024 – Jan 2025 & Dec 2025 – Jan 2026
  • Facilitated lab sessions on Graph Neural Networks (GNNs) created on top of lectures delivered by Prof. Michael Bronstein and team, Large Language Models (LLMs), and Vision-Language Models (VLMs)
Instructor
AI-Assisted Disease Screening Bootcamp, NAAMII · Aug 2025 – Oct 2025
  • Designed and delivered a curriculum combining lectures, paper discussions, and a week-long project phase; trained participants to build AI-based tools for oral cancer screening
Freelance Developer
Upwork (Remote) · Nov 2023 – Aug 2024
  • Built a Python application for automated news extraction and structured display from skynews.au
  • Developed a data scraper for a real estate listings platform

Research & Development

Academic research projects and practical development work

Research Projects

🤖
AI-Generated Image Detection
Final Year Thesis Project

AI-Generated Image Detection

Implemented FreqNet deepfake detection model using convolutional layers in the frequency domain (fast Fourier transforms); investigated performance against CLIP — best model achieved AUC 0.97. Found critical failure modes when fusing frequency and spatial domain features. Evaluated across 10 datasets and 6 deepfake generation architectures.

PythonPyTorchNumPyOpenCV
👆
Cross-Platform Fingerprint Matching System
Third Year Minor Project

Cross-Platform Fingerprint Matching System

Trained a Siamese network (VGG-16 + DenseNet) for cross-platform fingerprint verification; achieved 89% accuracy.

PythonTensorFlowNumPyPillow
🧬
Breast Cancer Classification — Hybrid Classifier
Solo Project

Breast Cancer Classification — Hybrid Classifier

Designed a custom hybrid architecture combining ViT, LBP, and EfficientNet for breast cancer classification.

PythonPyTorchScikit-learnPillow
🖥️
DevOps-Centric Image Classifier
Course Project

DevOps-Centric Image Classifier

Built a full-stack image classification web app (ResNet backbone) with a Node.js frontend and FastAPI backend; containerised and orchestrated via Docker + Kubernetes.

PythonPyTorchFastAPINode.jsDockerKubernetes
🏺
3D Reconstruction using Gaussian Splatting
Advanced 3D reconstruction technique for object modeling

3D Reconstruction using Gaussian Splatting

Implemented state-of-the-art 3D reconstruction using Gaussian Splatting technique for high-quality vase modeling. Created interactive 3D visualization demonstrating advanced computer vision and 3D modeling capabilities.

Computer Vision3D ReconstructionGaussian Splatting

Development Projects

📰
Automated News Extraction System
Python application for real-time news aggregation

Automated News Extraction System

Built a comprehensive Python application for automatic news extraction and display from skynews.au. Features real-time data processing, content filtering, and user-friendly interface for news consumption.

PythonWeb ScrapingBeautifulSoupAutomation
🏠
Real Estate Data Scraper
Comprehensive property data extraction tool

Real Estate Data Scraper

Developed a robust data scraping solution for real estate websites. Implemented advanced parsing techniques, data validation, and export functionality for market analysis and research purposes.

PythonData MiningSeleniumData Analysis

Technical Skills

ML / DL

Python, PyTorch, TensorFlow, Scikit-learn, NumPy, Pandas, OpenCV, Pillow

LLMs & GenAI

Transformer architectures, PEFT / LoRA, RLHF, prompt tuning, LLM evaluation

Engineering

FastAPI, Node.js, Docker, Kubernetes, Git, Unix/Linux, MySQL

Soft Skills

Clinical Collaboration, Technical Writing, Research Communication, Teaching

Certifications

AI Course — Samsung Innovation Campus (SIC)
Samsung (Onsite in SIC, Nepal) · 6 months
ML (classification, clustering, anomaly detection), NLP, CNN/RNN/GANs, dimensionality reduction, language models.
View Certificate ↗
Deep Learning Specialization
Coursera — Andrew Ng / DeepLearning.AI · 115 hrs
Neural networks, CNNs, RNNs/LSTMs/GRUs, attention mechanisms, hyperparameter tuning, regularisation, transfer learning, sequence models, structuring ML projects.
View Certificate ↗
Generative AI with Large Language Models
AWS / DeepLearning.AI · 16 hrs
Transformer architecture, PEFT/LoRA, soft prompts, RLHF, multi-GPU training strategies, LLM evaluation benchmarks.

Research Interests

Self-supervised Learning Vision-Language Models World Models Neural Cellular Automata Domain Adaptation Medical Imaging Multimodal Representation Learning Active Learning Learning with Label Noise

Gradient Descent: The Game

You are the model. Answer deep learning trivia correctly → gradient step toward minimum. Wrong → gradient ascent.
Loss L(θ) —
Steps 0
Streak ✓ 0
You are the model. Answer correctly → gradient descent step. Answer wrong → gradient ascent. Set η and press ▶ Start.