I have built several end‑to‑end AI systems that combine data engineering, modeling, and user‑facing interfaces. Below are the three flagship projects that showcase my full‑stack machine‑learning and agentic AI expertise.
AI Energy Recommendation Engine – Developed a full‑stack machine learning application that analyzes household energy consumption patterns and generates personalized recommendations across cost, safety and efficiency categories. Built a classification pipeline on a 2M+ record dataset (UCI), applying feature engineering to transform raw consumption readings into meaningful behavioral signals, and trained a Random Forest model achieving a 73% F1-score. Served predictions through a FastAPI backend and built a React frontend for end users to view personalized, category‑specific recommendations in real time. Presented the system and its design rationale in a final academic evaluation to external examiners.
Multistep Agentic AI System – Architected and built a multi‑agent orchestration system using a 3‑agent pipeline (Retriever, Analyzer, Writer) coordinated through FastAPI, with Redis pub/sub for inter‑agent messaging and Server‑Sent Events (SSE) for real‑time streaming of agent state and outputs to the client. Designed the system to support asynchronous, non‑blocking agent execution where each agent processes and hands off independently rather than running as a single monolithic LLM call and achieving 85% correctness across 1,000+ test evaluations. Used Pydantic for strict schema validation across every agent‑to‑agent handoff, ensuring reliable structured outputs through the full pipeline.
RAG QA System – Designed and engineered a production‑grade Retrieval‑Augmented Generation platform from the ground up, combining hybrid retrieval (BM25 keyword search + dense vector search via Qdrant) with FlashRank cross‑encoder reranking to maximize retrieval precision. Implemented a LangGraph‑based agentic retry loop that allows the system to self‑correct and re‑query when initial retrieval confidence is low, rather than returning a single‑shot answer. Built table‑aware document parsing with pdfplumber to correctly extract and preserve structured data (tables, layouts) that standard chunking pipelines typically destroy. Evaluated system quality using the Ragas framework, achieving a Faithfulness score of 1.0 on successfully judged questions, and instrumented full pipeline observability with LangSmith for tracing and debugging. Delivered via a Streamlit frontend for interactive querying.
These projects demonstrate my ability to move from raw data to deployed AI products, integrating backend services, front‑end interfaces, and rigorous evaluation metrics. For more background on my academic training, see my education.