Selected ProjectsSelected Projects

AI-native full-stack app with eight integrated modules covering the full end-to-end workflow -- CRM, proposals, projects, timesheets, invoicing, expenses, contracts, and reports. One data model connects leads to invoices to financial reports. Eight Claude Code skill files let an AI agent operate the entire system through natural language.

End-to-end self-serve AI platform with 15+ LangGraph agents that automate data ingestion, transformation, ML model training, and on-demand analytics. Features a natural language interface that converts questions into SQL, charts, and narrative explanations through a multi-agent decision routing system.

Enterprise SaaS for AI-powered document intelligence with dual-mode adaptive RAG. Agentic LangGraph pipelines mimic how human analysts work: search for content, flip to specific pages for visual analysis, reflect on coverage, then extract with confidence scores and source citations.

Stateless middleware that protects LLM applications with a progressive 5-layer input guard and 4-layer output guard. Uses a fast-first architecture: regex at ~2ms catches obvious attacks, BERT at ~150ms handles semantic threats, and LLM judge at ~2s resolves edge cases -- keeping median latency low.

Income estimation model (MAPE ~25%) and loan screening model (AUC 0.75) built on millions of national credit records. Includes a 6TB PostgreSQL data warehouse with daily ETL replication and an internal MLOps platform prototype with automated feature selection.

Two-tier email relay system replacing 60+ manually-rotated Gmail accounts for a debt collection agency. Tier A uses Postfix with a custom Python round-robin daemon for automatic Gmail account rotation. Accompanied by a reusable 8-role Ansible VPS provisioning toolkit with security hardening.