Sivangi Chatterjee

PennCloud

Distributed Systems

Built a fault-tolerant distributed cloud platform in C++ featuring webmail and file storage across 5 replicated backend servers, supporting file uploads of 800MB+. Implemented primary-based replication across 10 tablets (each on 3 replicas), write-ahead logging, and checkpointing with two-path crash recovery ensuring zero data loss across node failures. Includes a master coordinator with automatic primary election within one 250ms heartbeat cycle, a frontend load balancer, and an admin console for live fault injection and recovery testing.

Machine Learning Research

Studied how knowledge distillation reshapes failure modes in Retrieval-Augmented Generation systems. Built a controlled RAG pipeline on the SciFact biomedical benchmark, evaluated six chunking strategies, and distilled a 1B-parameter student from a 70B teacher using LoRA fine-tuning. Categorized failures using the Seven Failure Points taxonomy across 338 validation claims. While accuracy dropped 14.2 points, the student hallucinated roughly 30x more often (0.3% → 8.9%) and amplified the teacher's label bias, demonstrating that aggregate accuracy hides critical deployment risks in distilled RAG.

Generative Social Persona Modeling and Simulation

Machine Learning Research

Built a generative simulation framework to tackle extreme class imbalance in social data. The system synthesizes realistic responses across 25 distinct personas, enabling model training on heavily skewed datasets. I benchmarked classical and neural approaches before fine-tuning a language model with LoRA, achieving strong gains over zero-shot baselines.

Bidirectional Co-Attention for Image-Text Retrieval

Deep Learning Research

Designed a cross-modal retrieval architecture that aligns visual and textual representations through bidirectional co-attention. By fusing frozen ViT and BERT features before pooling, the model captures fine-grained correspondences between image regions and text tokens, outperforming standard contrastive approaches on Flickr30k retrieval benchmarks.

Job Market Analytics and Intelligence Engine

Full Stack Development

Created an end-to-end analytics platform that aggregates and visualizes hiring trends across a large-scale job market dataset. The system features a carefully normalized database schema, RESTful APIs, and an interactive frontend for exploring hiring hotspots, with query optimizations that brought analytics response times well under a second.