projects

Independent Research Projects

Develop personalized AI-driven food recommendation system to improve mood through dietary interventions. Built dataset to track and analyze food, exercise, sleep, physiology, circadian rhythm and mental health state.

Machine Learning & AI Projects

Implemented and evaluated image segmentation pipeline using TensorFlow and Segment Anything (SAM). Achieved IoU=0.31 for optic disk and IoU=0.08 for fovea segmentation for automated retinal anomaly detection.

Classified presidential candidate speeches via supervised and semi-supervised learning in Python/TensorFlow. Built n-gram language models on the Brown, Gutenberg and Reuters corpuses. Analyzed in-domain and out-of-domain perplexities to compare language models and individual sentences. Developed a part-of-speech (POS) and named entity recognition (NER) tagger for twitter data using Conditional Random Fields (CRF) and incorporated Viterbi algorithm to improve CRF accuracy. Implemented top-K sampling, nucleus sampling, beam search decoding algorithms and evaluated summarization models qualitatively and quantitatively using Python/TensorFlow.

Implemented full-shot and zero-shot dialogue state tracking on MultiWoz 2.1 dataset with 5 domains and 8438 dialogues using Python/TensorFlow to transfer knowledge from resource rich domains to unknown domains Deployed BERT base model and evaluated accuracy for inserting slot descriptions in zero-shot and full-shot DST

Classified fashion-MNIST dataset running convolutional neural networks (CNN) on Google Colab using Python Achieved 95.88% training accuracy and 93% test accuracy after hyperparameter tuning and cross-validation

Programmed reinforcement learning agent using Monte Carlo Tree Search in Python to solve Sokoban puzzle Designed and implemented machine learning algorithms using kNN, Naïve Bayes classifiers, linear regression, cross-validation, logistic regression, shattering, nearest neighbor, decision trees, neural networks, and clustering

Web Development & Data Science Projects

Crawled 50,000 URLs from ics.uci.edu domain using Python to find page similarity and subdomains Built search engine using Flask, HTML, CSS to query and retrieve top twenty matches from crawled databases

Developed first-of-its-kind energy parameter visualization platform for 200+ countries using Dash Deployed scalable and globally accessible website using Heroku sourcing data from a structured SQL database using SQLite Actualized user-friendly interface for parameters with customizable checkboxes and predictions using logistic regression in Python

Embedded Systems & Hardware Projects

Achieved average 85% throughput for FIR filter, DFT, FFT using Vivado High Level Synthesis (HLS) Added a new benchmark to Spector HLS, a benchmark suite for FPGA by implementing canonized Huffman Encoding in C++ Optimized design space with 15% higher throughput range and 60% greater pareto points compared to baseline

Outperformed traditional sensing techniques with remote soil sensing and active real-time pest deterrence using Linux, C/C++ Introduced predictive capabilities within 10% sensing range based on linear regression using the Scikit-learn library in Python Visualized soil vitals on an interactive online dashboard developed using HTML, CSS, Flask, and JavaScript

Attained 70% accuracy in determining an unknown amount of grocery waste using C/C++ and principles of RF attenuation Observed less than 25% standard deviation during prototype testing using received signal strength indicator (RSSI) metric Realized hands-off food waste estimation without modifying existing trash bin structure by simple retrofitted add-ons

Developed wireless temperature sensing framework using a resistance temperature detector (RTD) sensor with less than 0.2 variation between sensed and actual values Achieved 20% less external noise interference using a Sallen-Key low-pass filter in read-out circuit built using PSoC creator Executed real-time secure communication with less than 5% latency using C/C++ with data encapsulation and visualization