Pramod Pasala

• Built a Streamlit analytics dashboard that identified faulty hydraulic cylinders at customer sites using position-diff data and trend visualizations
• Developed and optimized time-series anomaly detection models for a predictive maintenance platform
• Provisioned AWS infrastructure using CDK: an ECS-based training cluster, Docker images for model training, and a local-trigger SageMaker estimator
• Investigated root causes of ML model failures in production deployments
• Supervisor: Vishnu Valsalan, Senior Data Scientist. Zero absences.

• Wrote Python pipelines for data validation and anomaly detection across large JSON datasets
• Built desktop automation tools with OpenCV and spaCy for UI inspection
• Used SPARQL to extract structured data from RDF knowledge graphs
• Created data comparison frameworks across software versions

Designing a framework that constrains LLMs to reason only over verified knowledge graph paths, reducing hallucinated responses in supply chain audit queries. The approach forces models to ground their reasoning in a Neo4j knowledge graph rather than generating freely.

• Evaluating 6 LLM models (3B to 122B parameters) across 5 reasoning paradigms: Naive LLM, Vector RAG, GraphRAG, GCR, and GCR + Graph-Augmented Chain-of-Thought
• Building a Neo4j knowledge graph with 8 node types and 8 relationship types
• Implementing prompt engineering patterns: enum-constrained JSON schemas, forced abstention, repair loops
• Preliminary results: GCR significantly reduces hallucinations across all models tested

Coursework: AI & Software Development, Advanced Machine Learning, NLP, Computer Vision, Feature Engineering, Information Theory, Data Storage & Analytics

Grade: 87.61%