An open-source full-stack platform that unifies grant discovery and researcher matching across a corpus of 80,000+ U.S. federal grants. Its semantic retrieval engine (Cohere embed-v4 embeddings on S3 Vectors) outperformed BM25, TF-IDF, and keyword baselines (MRR 0.74 vs. 0.54; nDCG@5 +47%) over 500 researcher profiles. Grants.gov ingestion is automated via AWS Glue ETL, deployed serverless on Lambda with Cognito auth and GitHub Actions CI/CD.

A containerized Kafka + Spark Structured Streaming pipeline (Docker Compose) that ingests NYC TLC taxi trips and computes 10-minute zone aggregates in PostgreSQL. A memory-aware XGBoost fare predictor (continued boosting over batched parquet) reaches 3.42% MAPE — a 78% error reduction over the zone-hour-mean baseline (MAE $0.49 vs. $1.99) — surfaced in a live analytics dashboard.

A privacy-first, fully local RAG system that ingests academic PDFs into a semantic vector store and synthesizes grounded outputs via a locally-served LLM (Llama 3 / gpt-oss 20B). Built with LangChain, ChromaDB, and HuggingFace MiniLM embeddings on Ollama, it eliminates external API calls and cloud costs entirely.

A Fashion-MNIST clothing classifier built with softmax and One-vs-Rest logistic regression over PCA features, reaching ~85% test accuracy while PCA cuts training time ~3.5x; rotation robustness was analyzed as part of the evaluation. Served end-to-end via FastAPI + React with real-time camera inference and background removal.

A hybrid quantum-classical Transformer that replaces the final encoder block's FFN with a parameterized quantum circuit (data re-uploading + entanglement) in PyTorch and PennyLane, trained end-to-end via the parameter-shift rule. It matched a classical Transformer baseline (R² 0.88 vs. 0.90) with ~6% fewer parameters and retained R² 0.86 under IBM Quantum hardware-noise evaluation.