I build analytical and AI systems from initial data preparation through evaluation and delivery. My public work includes customer segmentation and forecasting pipelines, churn models and Power BI dashboards, document-grounded RAG, structured LLM workflows, and FastAPI services backed by relational and document databases.

A FastAPI inventory system that separates transactional operations in PostgreSQL from audit and conversation history in MongoDB.

• Seeded and managed 22 products, 4 warehouses, 88 inventory records, and 320 movement records.
• Implemented JWT authentication, request tracing, FIFO stock movement, low-stock detection, and immutable audit logs.
• Exposed 12 permission-aware agent tools for querying and modifying inventory through natural language.
• Tested more than 20 agent queries and documented matching, error-handling, and response-formatting limitations.

A LangGraph workflow for extracting resume and job requirements, analyzing skill gaps, and returning structured candidate–job fit evidence.

• Parallelized resume and job-description extraction with asynchronous LangGraph nodes.
• Reduced end-to-end Gemini pipeline latency by approximately 30–40%, bringing tested runs below 40 seconds.
• Added typed Pydantic outputs and documented failure modes involving coursework inflation, domain pollution, and invalid transferable-skill mappings.

A document-grounded question-answering application for AI and machine-learning research papers.

• Built PDF ingestion, chunking, 384-dimensional MiniLM embeddings, and ChromaDB HNSW retrieval.
• Connected retrieved context to gpt-oss-20b through Groq and exposed the workflow through FastAPI.
• Containerized the application with Docker and documented delivery through GitHub Actions, AWS ECR, and EC2.
• Added a benchmarking harness for comparing ANN implementations by build time, memory, query speed, and Recall@K.

An analytical pipeline for customer behavior, product relationships, and revenue forecasting.

• Cleaned 541,909 retail records to 398,139 validated transactions by removing cancellations, missing customer identifiers, invalid prices, and administrative entries.
• Identified four customer segments using RFM features, K-means clustering, and silhouette-based model selection.
• Generated association rules for product bundling and inventory-planning opportunities.
• Selected a SARIMA specification with test RMSE of approximately 25,108, compared with approximately 28,060 for the initial specification.

Two complementary projects examining the same business problem from reporting and predictive-modeling perspectives.

• Customer Churn Analysis Dashboard: analyzed 7,043 customers in Power BI, measuring a 27% churn rate and approximately $139.13K in monthly revenue exposure.
• Telecom Churn Prediction: trained a Keras ANN that reached 91.22% validation accuracy and 0.81 F1 for the churn class.
• Examined contract type, tenure, payment method, service bundles, billing, and support activity as retention indicators.

Across these projects, I focus on:

• validating and documenting data-quality decisions before modeling;
• selecting evaluation methods that match the problem, including class-specific metrics, time-based testing, and leakage-aware validation;
• packaging models and analytical workflows so they can be reproduced or used through an application;
• documenting limitations, failure modes, and operational trade-offs alongside successful results.