AI study companion that turns PDF notes into summaries, quizzes, flashcards, and a memory-aware chat assistant.

• Built map-reduce summarization, auto-generated quizzes with SQLite-tracked score history, and flashcards scheduled via SM-2 spaced repetition
• Implemented history-aware RAG chat using LangChain + ChromaDB + Groq, with answers grounded in source page numbers
• Designed persistent vector storage — documents are embedded once, so re-uploading the same PDF reuses the existing ChromaDB index instead of re-embedding
• Evaluated every pipeline with RAGAS + LLM-as-Judge (chat faithfulness 0.87–0.94, quiz/flashcard/summary scores 4.0–5.0/5), with regression checks automated in GitHub Actions
• Containerized with Docker, deployed live on Render

End-to-end ML & MLOps system that predicts whether a patient will miss a scheduled medical appointment, built on 110,527 real hospital appointment records.

• Benchmarked 8 classifiers and selected a tuned LightGBM model, prioritizing recall (0.814) and ROC-AUC (0.737) over raw accuracy — since missing a likely no-show is costlier than a false alarm
• Built the full pipeline: data validation → EDA → feature engineering → MLflow experiment tracking & model registry → SHAP explainability
• Shipped a FastAPI prediction service with 29 automated tests, Dockerized and deployed live on Render, with GitHub Actions CI running tests and builds on every push

Agentic AI workspace for understanding and analyzing software projects, built on the Model Context Protocol (MCP) — three specialized MCP servers (Filesystem, Git, Knowledge/RAG) orchestrated by a single LangGraph agent.

• Built Filesystem, Git, and Knowledge (RAG) MCP servers with FastMCP, each exposing tools the agent can call — file exploration, commit/diff analysis, and FAISS-backed semantic doc retrieval
• Implemented a LangGraph agent (Groq gpt-oss-120b) that dynamically selects and chains tools across servers, grounding every response in actual tool output rather than invented information
• Added security boundaries (path sandboxing, sensitive-file blocking, bounded Git inputs) and a similarity-threshold gate on retrieval so low-confidence matches are discarded, not surfaced
• Supports analyzing the current project, a local repo, or cloning and analyzing any public GitHub repo on the fly
• Validated with 21/21 passing MCP server tests and a 26-scenario agent evaluation suite (92.3% pass rate); Dockerized and deployed live on Streamlit Community Cloud

Predictive maintenance pipeline for Diesel Particulate Filters (DPF) on commercial vehicles — predicts soot load from live engine and telemetry data to flag regeneration needs before failure.

• Generated and engineered features from 172,800 rows of synthetic sensor telemetry, merged with maintenance and trip-characteristic data into a 35-feature pipeline, with rolling 1h/3h windows and out-of-range/missing-value handling
• Benchmarked Random Forest vs. XGBoost, with Random Forest reaching R² 0.9997 (RMSE 0.601%, MAE 0.307%) on a held-out 34,560-row test set; all runs tracked via MLflow
• Served predictions through a FastAPI service with 4 endpoints — single prediction, batch/fleet-wide prediction, model info, and health check — with interactive Swagger docs
• Covered with an 11-test pytest suite, Dockerized, and validated on every push via GitHub Actions CI

• Trained 4 classifiers (Random Forest, SVC, AdaBoost, Decision Tree) on LAPD crime data — Random Forest was the best performer
• Engineered time-based, geographic, and demographic features; applied GridSearchCV hyperparameter tuning
• Conducted feature importance analysis to identify key predictors across location, timing, and incident type

• Built an end-to-end ML observability pipeline with SHAP explainability across demographic groups and data drift detection (KS, PSI tests) triggering automated retraining alerts
• Automated the full pipeline (train → evaluate → explain → monitor) via GitHub Actions CI, cutting manual intervention to zero
• Integrated MLflow for experiment tracking, model registry, and reproducibility

A production-deployed REST API for managing students, courses, enrollments, and grades with JWT-based authentication.

• Built full CRUD routes for Students, Courses, Enrollments, and Grades using FastAPI with auto-generated Swagger docs at /docs
• Implemented JWT authentication (python-jose + passlib/bcrypt) — all routes are protected and require a valid token
• Used SQLModel (Pydantic + SQLAlchemy) for ORM, with duplicate checks for emails, course codes, and enrollments; marks validated between 0–100
• Deployed live on Railway with a Procfile + Uvicorn ASGI server

• Analyzed sales and inventory data for a retail mall (~₹2,00,000/month revenue) to identify operational gaps
• Identified top product categories contributing ~70% of total sales via Pareto analysis; surfaced seasonal trends (November peak, January low)
• Delivered 6 actionable recommendations projected to reduce capital blockage by 20–25%

• Built a multi-role quiz platform (V1) with admin and user roles, quiz creation, scoring, and result analytics
• Extended to V2 with REST API, Vue.js frontend, Celery background jobs, and interactive performance dashboards