Machine Learning Researchist
Robust background in Software Engineering and MLOps.
Eager to contribute to cutting-edge ML research.
Feel free to connect on my social networks, collaborate or just geek out over ML!
Open for Research collaboration
Machine Learning
Artificial Intelligence
Skills
Jobs & Research
- Silicon content prediction via ML in blast furnace
- 90% Accuracy
- Research paper accepted
- Energy efficiency because of good silicon predictability implies immense economic impact.
- Hearth erosion modeling
- Implemented first principle heat transfer models (forward)
- Engineered optimization flow (inverse process) via tuning, choosing step sizes, making manifolds smooth and choosing the best optimization algorithms
- ML state models and ROMs used fasten computation from 5-10 minutes per 1 forward iteration to 10 seconds
- Research paper underway
NLP & Efficient LLM research with a publication in EMNLP 2024
- Computer Vision & Attention Mechanisms in Medical Imaging and Live Surgical Feeds.
- Research Published in IEEE
- Improved image localization & detection models for accurate market penetration count - 90% accuracy & 70,000 USD savings
- Developed image detection models, OCR & clustering models for real-time license plate number detection - prompt bonus delivery and minimal manual intervention with associated costs
- Established data pipeline, modeling and deployment flow for risk assessment models to predict fraud - prototype deployed (0.95 F1 score) that handled tens of thousands of transactions daily.
- Received the highest bonus in the year (4 people among 150 in the office)
- Developed LSTM and Conv1D models for breathing waveform pattern classification. - improved final EIT lung diagnosis
- Developed & deployed EIT Amplitude Image Classifier using two distinct iterative improvement approaches - improved final EIT diagnosis.
- Developed & deployed Gense Mobile App
- User account Management
- Live Test sync with Backend Server
- Test results storage & visualization
- Published a research paper in IEEE
- Worked on User Input UI of a charging system
- Produced maintainable, reusable and and portable code using Electron & React for:
- Web
- Desktop
- Mobile (IOS & Android)
- Refactored microservice responsible for grouping orders, reading client inputs
and forming routes - Microservice in various countries and cities like Malaysia, India & Bangkok
Publications & Preprint
- Forecast silicon content with 90 % accuracy owing to:
- Comprehensive data processing pipeline for large scale industrial data
- Hyperparameter optimization
- Model Selection involving NN, Xgboost & Time series Models
Blast Furnace
Silicon Content
Xgboost
Time Series Models
- Pipeline for efficient LLM training for:
- Reduced Miscalibration -10%
- Improved Accuracy +2.3%
- Distillation with trustworthy maximization process
- Substantially reduced training data size
Efficient LLM
Llama
NLP
- Improved smoke removal in live surgical feeds using sequence of frames
- Separate Temporal Mechanisms compared in CycleGAN base skeleton:
- Attention
- Convolution
- LSTM
- • Better smoke removal than benchmark DeSmoke CycleGAN model based on the JNBM & FADE metrics.
Computer Vision
CycleGAN
Attention Mechanism
Medical Imaging
- Modified CycleGAN archictecture
- Converted low resolution EIT images to high resolution CT images
- Mutual Information loss construction yielded structurally aligned generated CT images
- Normalized Mutual Information (NMI) gain from 0.2600 to 0.2621, (p<0.0001) from vanilla approach
Computer Vision
CycleGAN
Mutual Information
Medical Imaging
Healthcare Diagnostics
- An extension journal paper of IEEE EMBC paper
- Incorporates modified CycleGAN to infer prior EIT image
- Dynamic Prior yields a better final reconstructed EIT image
- NMI gain from 0.45 to 0.49 from vanilla approach
Computer Vision
CycleGAN
Medical Imaging
Surgical Blender: A Synthetic Datat Generator for Robot-Assisted Surgery
Contribution: Second Author
Status: Results pending in CBM Journal
- Experimented Improved metrics for CycleGAN using
- Carefully crafted synthetic data
- Custom IC & DC Losses
- Potential to alleviate the load for acquiring real medical images in training tasks like smoke removal
Computer Vision
CycleGAN
Medical Imaging
Education
• 4.0 GPA thus far ¯_(ツ)_/¯
• First Class Honors
• Deans Honor List in 2016-2017 & 2019-2020
• HKU Foundation Scholarship 2016-2020 (for outstanding undergrads)
• Young Tsun Dart Scholarship 2017 (reserved for only one student in a particular year of study)
