Developed ChatBot using Streamlit as frontend, and retreival-augmented generation with GPT-3.5-turbo and vectorized private data using ChromaDB as backend.
Paper Category Classification
Prompt engineered GPT-3.5-turbo to classify the category of a medical research paper using the "abstract" section as input. Achieved 20% increase in F1-Score.
Pill Classification
Developed pill classification model using a multi-head metric learning algorithm. Achieved classification rate of 90%(classified 10% as unknown) with 100% accuracy on 296 classes.
Model Monitoring Platform
Developed a plotform to monitor model performance. Used Nvidia-Triton Server for model deployment, K6 for performance testing, Prometheus for storing time-series data, and Graphana for metric visualization.
Model Retraining UI
Developed an UI for non-developers using jupyter notebooks. The user can decide training configurations and generate configuration files from those settings for reproducibility without any code.
Section Angle Detection
Developed a semantic segmentation model to detect the section of an object, and implemented a deterministic algorithm to calculate the angle of a certain point within the outline of the section that is specified by the user. Created UI using OpenCV. Achieved 92% cut in time cost.
Data Preparation Platform
Developed an platform where you can import data from multiple sources with multiple options for preprocessing without any code. Used NextJS as frontend, Django as backend.
Water-Level Detection on Edge Devices
Developed a water-level detection app which runs inference on images from surveillance cameras using Nvidia-Deepstream.