BISHWASH KHANAL

AI Engineer & Computer Vision Researcher
khanal.bishwash08@gmail.com
Jyväskylä, Finland

Education

Master's Degree Programme (M.Sc.) in Artificial Intelligence
University of Jyväskylä
Jyväskylä, Finland
08.2024 - present
Relevant Courses: Deep Learning for Cognitive Computing, Semantic Web and Linked Data, SOA and Cloud Computing, Collective Intelligence and Agent Technology, Computer Vision and Image Analysis, Natural Language Processing, and Simulation.
Bachelor of Science in Electrical and Computer Engineering
Jacobs University Bremen
Bremen, Germany
09.2017 - 08.2020
Minor in Intelligent Mobile Systems; Grade: 1.73 (German grading system)
Relevant Courses: Signal Processing, Wireless Communication, Information Theory, Machine Learning, Computer Vision, Artificial Intelligence, Robotics, Embedded Systems.
Thesis: Non-Linearity in Wireless Communications and Deep Learning

Professional Experience

Artificial Intelligence Engineer
OptiML Org (part-time, remote)
California, USA
Mar 2024 - present
LLM Finetuning and Compression (2024-present)
This project involves integrating powerful model compression techniques into fine-tuning workflows, empowering AI developers and researchers to optimize and deploy large models with efficiency and scale. Founding member of the team.
My Responsibilities: Develop a custom pipeline for fine-tuning and compressing open-source LLM models together, evaluating their performance on domain-specific tasks; research and evaluation of SOTA LLM Compression techniques.
Tools Used:
Python (torch, transformers, datasets, peft, bitsandbytes, accelerate, wandb)
Models Used: Llama, Qwen, Mistral, Gemma, Phi, Deepseek
Computer Vision Engineer
E.K. Solutions Pvt. Ltd. (full-time, on-site)
Lalitpur, Nepal
Mar 2021 - Aug 2024
Model Compression (2024)
This project involved the quantization of computer vision models for face detection, recognition, and swapping to enable real-time performance on edge devices like the Jetson AGX Orin.
My Responsibilities: Set up and deploy face detection, recognition, and swapping models on Jetson AGX Orin; perform int8 model quantization using PyTorch Quantization and NVIDIA TensorRT for optimized and real-time performance.
Tools Used:
Python (numpy, opencv, ultralytics, tensorrt, onnx, insightface)
Models Used: YOLOv8, Insightface (ARCFace, Inswapper), Hyperstyle (StyleGAN2, CLIP), SOLIDER
LLM Tools (2023-2024)
The project involved the development of a suite of multifaceted tools integrating Large Language Models (LLMs) for enhanced software project management, coding, documentation, meeting and video analysis, automation, QA, HR, and administrative tasks.
My Responsibilities: Development of a chatbot with a workspace containing user stories, transcriptions, and database analysis using Retrieval-Augmented Generation (RAG); integration of bulk Curriculum Vitae (CV) analysis.
Tools Used:
Python (langchain, openai, sqlalchemy, chromadb, fastapi, pandas)
Models Used: GPT-4 (through API), Sentence Transformers
Virtual Tours (2022-2024)
The project was about the generation of planar simplified 3D mesh for virtual tours of multi-room indoor scenes captured with iPad Pro Lidar, followed by texture projection using MVS texturing.
My Responsibilities: Geometrical analysis, formulation, and implementation of algorithms for planar simplification and texturing; implementation of ICP registration and TSDF volumetric integration for combining multiple 3D meshes; application of hole filling, mesh simplification, and texturing enhancement for 3D meshes.
Tools Used:
Python (numpy, opencv, open3d, fastapi, celery), Meshlab, C++, Swift5, Javascript (three.js)
Models Used: RandLA-Net (Semantic Segmentation)
Landmark Detection (2021)
The goal of the project was to develop a custom object detection model that detects faulty knot-bolt arrangements on large construction sites.
My Responsibilities: Labeling data for training deep Convolutional Neural Networks and YOLO models; fine-tuning pre-trained YOLO models based on new data and requirements.
Tools Used:
Python (numpy, tensorflow)
Models Used: YOLOv5, CNN
Facial Check-In System (2021)
This project involved the development of a system that would handle daily employee check-in/check-out through face detection.
My Responsibilities: Fine-tuning FaceNet models with an improved loss function; developing a pipeline for matching facial embeddings with the existing database using Naive Bayes Algorithm.
Tools Used:
Python (numpy, pytorch, scipy, scikit-learn)
Models Used: FaceNet
Competitive Coder and Data Analyst
Scale AI (freelancing, remote)
California, USA
06.2023 - 10.2023
Worked as a Freelancer through Upwork that involved data annotations to help train Generative AI models to become better writers.
My Responsibilities: Solving competitive programming problems to train AI model; ranking a series of responses that were produced by an AI model; writing a short story based on a provided topic; assessing whether a piece of text produced by an AI model is factually accurate or not.

Skills

Programming Languages
Python, C, C++, MATLAB, Java
ML Tools
Tensorflow, Keras, PyTorch, Scikit-Learn
ML Models
LLMs, Transformers, GANs, CNNs
CI/CD Tools
Docker, Kubernetes
API Frameworks
FlaskAPI, FastAPI, Jakarta
Database
MySQL, PostgreSQL
Environments
Linux, Unix, Windows
Miscellaneous Tools
Git, GitHub, LaTeX, VS Code

Publications

B. Khanal, M. Om, S. Rijal, V. Ojha (2024), Frontiers in Computer Science, Volume 6 - 2024, ISSN: 2624-9898
B. Khanal, S. Rijal, M. Awale, V. Ojha (2024), arXiv

Scholarships & Certifications

Scholarships

JYU Scholarship, University of Jyväskylä (2024-2026)
This scholarship is a university-level scholarship awarded by the JYU rector's decision awarded to students who demonstrate exceptional academic achievement. This scholarship covers 50% of the student's tuition fees.
Merit-Based Scholarship, Jacobs University Bremen (2017-2020)
This scholarship is awarded to students who demonstrate exceptional academic achievement, leadership potential, and a strong commitment to their field of study. This scholarship recognizes individuals who exhibit outstanding merit and promise in their academic pursuits.
Mahatma Gandhi Scholarship, Embassy of India in Kathmandu (2015-2017)
This scholarship is typically awarded to outstanding students who demonstrate exceptional academic achievements during High School and have a poor economic background.

Certifications

Notable Projects

Nepali OCR using Vision based Transformers (09.2023)
Inspired by the popular TR-OCR, exploring the possibilities of creating vision-based transformers for detecting Nepali text. The previous works have already been replicated and implemented on the datasets. A RoBERTa based tokenizer and encoder has been trained on Nepali datasets. Exploring on methodologies for collecting datasets for OCR.
Project Resources:
Nepali RoBERTa Model
Face Generator using Generative Models (03.2023-06.2023)
Explored and applied encoder-decoder frameworks, including autoencoders, VAE, DCGAN, and WGAN, to generate synthetic facial images with a focus on enhancing realism and diversity, utilizing the CelebA dataset. Conducted comprehensive mathematical and visual assessments across various model variants.
Google - American Sign Language Fingerspelling Recognition (05.2023-08.2023)
Participated in a competition organized by Google and hosted in Kaggle. Task included detection and translation American Sign Language (ASL) fingerspelling into text. Worked on a TensorFlow Lite model trained on labeled landmark data extracted using the MediaPipe Holistic Solution. The model achieved private score of 0.647 and public score of 0.696 which was able to secure 371th ranking out of 1315 teams.
Bremen Big Data Challenge 2020 (03.2020-04.2020)
Participated in the Big Data Challenge organized by Universität Bremen. The task included predicting the hand movements and actions performed with the help of sensor data (measured with Electromyography (EMG), bipolar derived). Used a combination of U-Net and LSTM models to get an error of 0.464 which was able to secure 12th out of 31 teams.