Cole McIntosh

Founder @Mellow AI

|

Engineer & Founder, Mellow AI · Dallas, TX

AI engineer and founder of Mellow AI. Builds production agents, document processing tools, and data products in Python, TypeScript, and Rust. Creator of Openextract, pdfmd, Liedar, Tradewind, and mev; contributor to LiteLLM, Pydantic AI, and smolagents.

Experience

Founder

2025 – Present

Mellow AI

  • Rebuilt the core extraction platform for a global law firm: new UI and architecture
  • Cut OCR cost 95% and latency 75% with a custom pipeline
  • Built openextract, open source LLM extraction in Python and TypeScript: 10+ model providers, v1.0 on PyPI
  • Trained mev, an open 4B decision model (LoRA on Qwen3.5-4B): 86.7% on the main dev set, calibrated confidence (ECE 0.038)

AI / Data Engineer

2023 – 2025

Pay Ready

  • Built a general support agent harness that prompts and skills steer into new workflows
  • Replaced a failing spam filter with a small LLM classifier (structured true/false) in the ticket workflow: ~60% less support agent work
  • Shipped a lease and contract extraction API on GPT-4 Vision: 250K+ documents processed; less than 1% required human correction
  • Built the Slack RAG assistant teams used daily instead of a bought tool: 60% less manual effort, 75% faster retrieval

Scaled Tooling Analyst

2022 – 2023

Otter / CloudKitchens (now Atoms)

  • Embedded with ops and engineering on internal tooling: Python pipelines scoring ~500GB/week
  • Shipped a Jira ticketing chatbot from discovery to production: 30% faster resolution

Business Analyst

2021 – 2022

RealPage

  • Built SQL models and dashboards for 75+ internal users; cut manual reporting 25%

Open source

  • Contributed provider integrations, streaming fixes, and usage reporting.

  • Added xAI search and attachments, input token counting, and provider settings.

  • Fixed Enum execution, added Bedrock API-key support, and improved agent tests.

Projects

  • Structured data from documents and media, validated against your schema.

  • A fast PDF-to-Markdown engine in Rust, built without dependencies.

  • Document screening that surfaces evidence of tampering for human review.

  • Public building permits turned into timely leads for DFW subcontractors.

  • I fine-tuned Qwen3.5-4B to choose one of 2–24 options from a structured state, and published the weights, training details, and evaluation.

Education & certifications

  • RAG, LangChain, LangGraph, Semantic Search, DeepLearning.AI
  • Azure Applied AI, Microsoft
  • AWS Cloud Technical Essentials, Amazon
  • Google Cloud Fundamentals, Google