A multi-agent RAG system that plans queries, retrieves and reranks passages, synthesizes answers with inline citations, and adds a reviewer critique. Wrapped in a PLP (personal learning portal) with modules, checklists, and notes.
Built a two-stage RAG stack from baseline to an advanced pipeline with query rewriting (Flan-T5) and cross-encoder reranking. Uses Milvus Lite for vector search.
Benchmarked GPT-5, Gemini Pro, and Gemma-3/4B on developer tasks (code understanding, API docs Q&A, troubleshooting). Compared CLEAR, few-shot, and chain-of-thought prompts for accuracy, coherence, and efficiency.
End-to-end QA pipeline: fine-tuning, packaging, parity checks, and Azure ML deployment.
Streaming ingestion → feature pipeline → MLflow tracking + Dockerization. Emphasis on experiment logging and model monitoring.
Hands-on Java project and labs exploring RPC, concurrency, replication, consistency, and sharded storage—plus design notes and tests.
Predicted churn for a bank’s 10k customers; built E2E pipeline from EDA → feature engineering → model selection. An SVM beat tree/boosting baselines and balanced precision with accuracy, surfacing drivers like credit score, age, balance, and satisfaction.