✍️ Writing: Shipping Enterprise RAG (4-part series)

Lessons from building a production RAG assistant for a regulated enterprise

A deep-dive series on what actually breaks in production RAG systems — and how to engineer around it. Drawn from shipping a RAG assistant over thousands of internal documents with real compliance stakes.

  • Part 1 — Stop optimizing your RAG prompt. The bug is two layers up. Why most wrong RAG answers are retrieval failures, not prompt failures.
  • Part 2 — Chunking strategies that beat the default (coming soon)
  • Part 3 — Hybrid retrieval: BM25 + vector + reranking (coming soon)
  • Part 4 — Production reality check: evals, observability, cost (coming soon)
  • Read Part 1 on this site Read on Substack & subscribe

Full-Stack Lead Assignment System

Production lead-routing platform — owned end to end, from data model to UI

A company-wide platform that automates how inbound leads are scored, routed, and tracked across sales teams — replacing a manual, error-prone hand-off process. I owned the full stack: the SQL Server data model, the FastAPI services, the business-rules engine, and the React dashboard the team uses every day.

The problem. Leads were distributed by hand, which meant slow response times, uneven workloads, and no visibility into where deals stalled. The business needed routing that was fast, fair, and configurable without a developer in the loop.

What I built.

  • A configurable business-rules engine so operations can change routing logic (territory, capacity, priority, round-robin) without code changes
  • Real-time dashboards for assignment status, team workload, and response-time analytics
  • A FastAPI backend exposing RESTful services over an MS SQL Server data model, with a React front end
  • Deployed and operated on Azure

Architecture: React (UI) → FastAPI (REST services + rules engine) → MS SQL Server, hosted on Azure.

  • Tech Stack: Python, FastAPI, React, JavaScript, MS SQL Server, RESTful APIs, Azure
  • Role: Sole full-stack engineer — data model, backend, rules engine, frontend, deployment
  • Impact: Automated lead distribution, faster and more even response times, and live visibility into the pipeline for the sales team

RAG Systems & SQL AI Agents

Intelligent AI systems for enhanced information retrieval and database queries

Developed production-ready RAG (Retrieval-Augmented Generation) systems and SQL AI agents that enable natural language interaction with databases and document repositories.

  • Tech Stack: Python, FastAPI, LangChain, LangGraph, Vector Databases, OpenAI API, MS SQL Server
  • Key Features: Semantic search, natural language to SQL translation, context-aware responses
  • Impact: Enabled non-technical users to query complex databases, improved information accessibility
  • Techniques: Embeddings, vector similarity search, prompt engineering

Spotify Music Recommendation API

Machine learning-powered music discovery and recommendation system

Built a comprehensive Spotify API wrapper with custom ML-based song recommendations. Implemented K-Nearest Neighbors algorithm to suggest similar songs based on audio features like danceability, energy, and acousticness.

  • Tech Stack: Python, Flask, Spotify API, scikit-learn, Heroku
  • Features: Track search, audio feature analysis, ML-powered recommendations
  • Results: Successfully deployed 3 production APIs serving recommendations
  • GitHub Blog Post

Climate Change Visualization Dashboard

Interactive global temperature visualization platform

Created an interactive web application to visualize historical climate data across countries and time periods. Features animated choropleth maps showing temperature changes from 1750 to 2013.

  • Tech Stack: Python, Plotly, Pandas, Flask, Heroku
  • Features: Interactive time-series animations, country-level analysis, Celsius/Fahrenheit conversion
  • Impact: Made complex climate data accessible and engaging
  • GitHub Blog Post

Heart Disease Prediction Model

Machine learning model with feature importance visualization

Developed a classification model to predict heart disease using patient health metrics. Implemented comprehensive feature importance analysis to identify key health indicators.

  • Tech Stack: Python, scikit-learn, XGBoost, Flask, Plotly
  • Results: Achieved 95%+ prediction accuracy
  • Features: Interactive web interface, real-time predictions, feature importance visualization
  • GitHub Medium Article

K-Nearest Neighbors from Scratch

Educational implementation of KNN algorithm

Built KNN classification algorithm from scratch using only NumPy, including custom StandardScaler implementation. Demonstrated on breast cancer dataset classification.

  • Tech Stack: Python, NumPy
  • Achievement: 97.6% accuracy on breast cancer classification
  • Purpose: Deep understanding of ML fundamentals, educational resource
  • GitHub Blog Post

Data Storytelling with Python

Visual analytics and exploratory data analysis

Series of data analysis projects using Miami-Dade County open data, focusing on meaningful insights and compelling data visualizations.

  • Tech Stack: Python, Pandas, Matplotlib, Seaborn, Plotly
  • Focus: EDA, statistical analysis, data visualization, storytelling
  • Medium Article

More Projects

Check out my GitHub profile for more projects and contributions, or read my blog posts for detailed technical writeups and tutorials.

Last updated: June 2026