Responsibilities include:

• Built an end-to-end pipeline to convert raw unstructured data into attributed graphs and benchmarked GNN architectures — GCN, GIN, and EdgeConv — for a regression task on sparse, high-dimensional targets.
• Designed a multimodal predictor combining graph embeddings, log embeddings, and state-space models to capture both structural and sequential signals from heterogeneous data sources.
• Developed a pre-training strategy using Triplet Loss to learn similarity structure from a proxy label before fine-tuning on sparse downstream targets, improving generalization on limited labeled data.
• Built a scalable data augmentation framework that programmatically generates and validates new graph instances, enabling robust dataset enrichment for ML benchmarking and robustness testing.
• Wrote an NLP-based feature extraction framework to parse unstructured logs into structured features for downstream ML modeling.
• Scaled GNN training to graphs with up to 300K nodes using distributed GPUs; built experiment tracking, dataset versioning, and containerized workflows for fast iteration across architectures.

Responsibilities include:

• Graph generation with geometrical and topological constraints. (Patent filed)

Responsibilities include:

• Deployed CNNs to predict 3D computer-aided design (CAD) models from 2D images.
• Automated segmentation and registration of point-cloud data obtained from scanning machine parts enabling efficient and accurate inspection.