02 — Projects
What I've Shipped
// PROJECT_01 · Saga Pattern & Event-Driven AI
⏳
Time-Traveling Agentic Orchestrator
Python · RabbitMQ · Redis · PostgreSQL · FastAPI
▸Challenge: Autonomous agents executing multi-step distributed transactions need guaranteed state reversibility and protection against race conditions.
▸Solution: Architected an event-driven AI orchestrator via RabbitMQ & FastAPI implementing the Saga Pattern for transactional safety. Engineered a PostgreSQL Shadow Ledger and Redis distributed locking to prevent race conditions and automatically execute reverse-chronological rollbacks on failure.
🔄 100% state reversibility via Shadow Ledger rollbacks
// PROJECT_02 · LangGraph & PostgreSQL
🧠
Agentic Text-to-SQL Engine with Autonomous Self-Correction
LangGraph · LLMs · PostgreSQL · Streamlit · FastAPI
▸Challenge: Naive text-to-SQL wrappers crash on complex queries — hallucinated schemas, syntax errors, and broken joins make them unreliable for production BI.
▸Solution: Engineered a stateful AI agent using LangGraph with an autonomous self-correction loop that parses PostgreSQL errors to dynamically rewrite and retry failed queries. Secured the microservice with database-layer Row-Level Security (RLS) for multi-tenant access control, exposed via REST API & Streamlit dashboard.
✅ 100% syntax validity before returning data to users
// PROJECT_03 · Enterprise AI Security
🛡️
Enterprise AI Gateway & PII Privacy Proxy
Python · FastAPI · Hugging Face (RoBERTa / Presidio) · Docker · Redis
▸Challenge: "Shadow AI" data leakage — employees paste proprietary code, customer PII, and financial data into cloud LLMs, violating GDPR and HIPAA compliance.
▸Solution: Built an AI middleware proxy that intercepts prompts via a local NER pipeline (RoBERTa/Presidio), masks PII before it leaves the network, sends sanitized prompts to the cloud LLM, then re-injects original data into the response — enabling complex reasoning without ever exposing protected data.
🔒 Zero PII exposure to external LLM APIs (HIPAA/GDPR compliant)
// PROJECT_04 · MLOps & Fine-Tuning
⚗️
Automated MLOps Pipeline for LLM Synthetic Distillation
Python · PyTorch · Hugging Face (PEFT / LoRA) · FastAPI · SQLite · Docker
▸Challenge: Running heavy foundation models (GPT-4o) for repetitive data extraction tasks is cost-prohibitive at scale — companies hemorrhage compute spend on tasks a smaller model could handle.
▸Solution: Designed an end-to-end MLOps pipeline that intercepts high-quality JSON outputs from a "Teacher" model into a local SQLite database, then automatically triggers LoRA/QLoRA fine-tuning on a 2B "Student" model once a 500-sample threshold is reached — enabling seamless endpoint swap from heavy to lightweight inference.
💰 90%+ reduction in token/compute costs while maintaining task accuracy