§ I

Selected Work

A handful of recent projects. A longer list lives on GitHub.

  1. IAI/ML
    LLM-Based Document Intelligence System

    Production AI system for automated document analysis, compliance verification, and structured report generation using advanced RAG architectures.

    Developed a comprehensive LLM-based system that processes unstructured documents, extracts key information, verifies compliance against regulatory frameworks, and generates structured reports. Implemented using LangChain with custom prompt engineering and RAG pipelines for enhanced accuracy.

    • LLMs
    • RAG
    • Document Analysis
    • Compliance
    • LangChain
  2. IIAI/ML
    RAG Pipeline with Vector Databases

    Engineered retrieval-augmented generation pipelines using Pinecone and ChromaDB for enhanced information retrieval from unstructured data.

    Built production-grade RAG systems that combine vector databases with LLMs to enable semantic search and context-aware responses. Implemented chunking strategies, embedding optimization, and hybrid search approaches for maximum retrieval accuracy.

    • RAG
    • Vector DBs
    • Pinecone
    • ChromaDB
    • Embeddings
  3. IIIComputer Vision
    Computer Vision Video Analysis Pipeline

    Developed CV pipelines for automated video annotation, object detection, and frame-level analysis using OpenCV and modern frameworks.

    Created end-to-end computer vision pipelines that process video streams, detect objects, annotate frames, and extract actionable insights. Implemented using OpenCV with custom detection models and optimized for real-time processing.

    • OpenCV
    • Object Detection
    • Video Analysis
    • Python
  4. IVBackend
    Secure FastAPI Backend for AI Services

    Built production FastAPI backends with comprehensive authentication, validation, logging, and error management for AI microservices.

    Engineered robust API backends that serve AI models with enterprise-grade security, including JWT authentication, request validation, structured logging, and comprehensive error handling. Deployed on GCP with auto-scaling and monitoring.

    • FastAPI
    • Python
    • Authentication
    • GCP
    • Microservices
  5. VAI/ML
    Agent-Based AI Workflow Automation

    Developed intelligent agent systems using LangFlow for USDA workflows with human-in-the-loop validation and controllability.

    Created agent-based AI tools that automate complex workflows while maintaining human oversight. Implemented using LangFlow with custom agents, tool calling, and validation checkpoints to ensure alignment with business requirements.

    • LangFlow
    • AI Agents
    • Automation
    • Workflow
  6. VISecurity
    OWASP LLM Security Implementation

    Implemented AI security best practices including prompt injection mitigation and OWASP LLM Top-10 compliance across production systems.

    Developed comprehensive security measures for LLM applications, including input validation, output filtering, prompt injection detection, and secure API design. Achieved full OWASP LLM Top-10 compliance for all deployed AI services.

    • AI Security
    • OWASP
    • Prompt Injection
    • Security
§ II

On Method

Every project I take on goes through the same three-beat rhythm: understand the problem in the customer’s own words, design the smallest system that can possibly answer it, then harden that system until it stays correct while I sleep.

The first beat is boring and the most important. I spend the most time here — in the user’s language, not the product’s. The second beat is where I resist every temptation to generalise: the simplest working system is always smaller than it first appears. The third beat is the one most engineers forget — a system isn’t shipped when it runs; it’s shipped when you stop worrying about it.

I write documentation as I build, not after. I prefer a plain deploy over a sophisticated one. I care more about how the third engineer feels using a codebase than how the first one felt writing it. None of this is opinion; it’s the result of shipping things that had to outlive the project page they were born on.