Selected work across continual learning, recommendation, computer vision, bioinformatics machine learning, and applied data systems.
Zillow Group
Continual user modelingDeveloped a mixture-of-experts architecture with rank-1, LoRA-style experts and a semantic expert-selection strategy for continual user modeling, outperforming the strongest baselines by 2.88% relative HitRate@5 on sequential recommendation and 1.37% accuracy on user-attribute prediction.
RIT
Continual learningDeveloped a continual structure adaptation framework that integrates beta-Bernoulli processes for structure inference within the sequential Bayes framework, enabling dynamic evolution of both network depth and width in continual learning scenarios.
RIT
Class-incremental learningDeveloped a cross-task representation-alignment framework that improved average accuracy by 2.16 percentage points for exemplar-free class-incremental learning.
RIT
Bioinformatics MLDeveloping a Bayesian adaptive graph neural network for gene-disease association prediction over protein-protein interaction graphs.
Pulchowk Campus
Computer visionDeveloped a license plate recognition system, tailored for Nepalese license plates, with three key stages: vehicle detection, license plate localization, and Nepalese character recognition.
Created a license plate localization dataset by annotating Nepalese vehicle images, alongside building a Nepalese character recognition dataset using Devanagari fonts.
Fusemachines
Multimodal MLLed the development of a multimodal machine-learning pipeline for identifying potential trafficking activity in online advertisements using video, image, and text data.
Built image–text contrastive models for advertisement matching, face-based identity linking, and BERT-based social-handle extraction, improving cross-ad linkage accuracy by 35%.
Fusemachines
Edge computer visionDeveloped a lightweight object detector that increased inference throughput by 47%, then deployed it on NVIDIA Jetson Nano devices for real-time waste-type and disposal-intent classification.
Fusemachines
Data analysisAnalyzed the impact of song quality, commercial length, and time of day on panelist retention, designing custom metrics and statistical tests to quantify song quality.
Resolved date-inconsistency bugs in the existing data pipeline and built a feature-engineering pipeline to augment the dataset.
Fusemachines
Anomaly detectionDesigned and validated an AutoEncoder-based semi-supervised learning pipeline for network anomaly detection, engineering session-level features from raw pcap traffic captures.