projects

Agentic RAG System

Hybrid Retrieval, Evals, Workflows

  • Built a production-style document Q&A pipeline handling PDF ingestion, chunking, and embedding into a durable, retry-safe workflow with full retrieval traceability.
  • Established a baseline retrieval precision of 68% by designing a ground-truth eval harness of 75+ Q&A pairs using exact-match and LLM-as-judge scoring.
  • Boosted eval pass rate to 89% and precision by 22% by implementing hybrid BM25 + cross-encoder re-ranking with agentic query-routing for low-confidence retrieval.

Python · Inngest · LlamaIndex · Qdrant · OpenAI

Secure Cloud File Management Platform

  • Enforced role-based access control across 8 endpoints by adding Owner/Editor/Viewer roles to a Spring Boot and React.js application.
  • Secured all endpoints with JWT-based Spring Security authentication, validated by 22 automated tests.
  • Cut API latency by 58% and raised throughput 75% by migrating large-file uploads to presigned S3 URLs, tested with k6 at 20 concurrent users.

Spring Boot · React.js · PostgreSQL · AWS S3 · JWT

Flight Delay Prediction System

  • Built a leakage-free ML pipeline on 5M+ flight records (BTS + NOAA, 15 hub airports) using point-in-time feature engineering; caught and fixed 2 real data-leakage bugs during development, backed by 9 regression tests.
  • Trained and compared XGBoost and PyTorch classifiers, improving PR-AUC 23% over a historical-rate baseline (0.320 vs 0.261) while identifying a ranking-vs-calibration tradeoff between the two models.
  • Deployed via FastAPI with MLflow experiment tracking, Docker containerization, and CI, serving both delay probability and expected delay-duration predictions.

Python · XGBoost · PyTorch · FastAPI · MLflow · Docker

F1 Race Prediction – Machine Learning Model

  • Reduced race-time error by 68% by building a pipeline evaluating scikit-learn and XGBoost models.
  • Accelerated deployment by 40% across 9 race experiments by tracking pipelines using MLflow and SHAP.

Python · TensorFlow

Chess Game

  • A terminal-based chess engine in modern C++20 with full move validation, check/checkmate/stalemate detection, en passant, castling, and promotion. Supports human vs human and human vs computer play across four difficulty levels, an optional X11 graphical board, undo, custom board setup, and a countdown clock mode.

C++20 · X11