# Cole McIntosh

**AI Engineer & Founder, Mellow AI**

[GitHub](https://github.com/colesmcintosh) · [LinkedIn](https://www.linkedin.com/in/colemcintosh/) · Dallas, TX · [cole@staymellow.ai](mailto:cole@staymellow.ai)

## About

AI engineer and founder of Mellow AI building production agent systems, document extraction pipelines, and developer tools in Python, TypeScript, and Rust. Combines hands-on systems work—from fine-tuned small models and zero-dependency parsers to multi-provider LLM infrastructure—with end-to-end delivery that measurably reduces operational load.

- **Role:** AI engineer; founder of Mellow AI
- **Based in:** Dallas, Texas, USA
- **Company:** Mellow AI (founded 2025), staymellow.ai
- **Focus:** Production AI agents, document extraction, model fine-tuning, and developer tooling
- **Languages:** Python, TypeScript, Rust, SQL
- **Notable work:** Core extraction platform for a global law firm (95% lower OCR cost, 75% lower latency); lease extraction API over 250K+ documents (<1% human correction)
- **Open source:** Merged PRs across LiteLLM (SSO, guardrails, OpenTelemetry, cost tracking, providers), Pydantic AI (xAI search/attachments, token counting), and Hugging Face smolagents
- **Contact:** cole@staymellow.ai

## Technical Stack

- **Languages:** Python, TypeScript, Rust, SQL
- **AI & Agents:** Agent harnesses, LLM evals & calibration, LoRA fine-tuning (Qwen), RAG, MCP, structured outputs
- **Frameworks & SDKs:** Pydantic AI, Vercel AI SDK, LangChain / LangGraph, LiteLLM, Zod, Next.js, Ratatui
- **Data & Cloud:** Custom OCR & PDF parsing, document forensics, OpenTelemetry & Langfuse, SQLite, AWS, Azure, GCP

## Experience

### Founder — Mellow AI
*2025 – Present*

- Re-architected the core document extraction platform and UI for a global law firm under strict data boundaries, modernizing ingestion and schema-validated pipelines.
- Engineered a custom OCR and extraction pipeline cutting processing cost by 95% and latency by 75%.
- Built and shipped Openextract (v1.0 on PyPI, plus a full TypeScript port)—a multimodal extraction toolkit supporting 10+ model providers with Pydantic/Zod validation.
- Fine-tuned and released mev (LoRA on Qwen3.5-4B), an open-weight structured decision model reaching 86.7% dev accuracy with calibrated confidence (ECE 0.038).

### AI & Data Engineer — Pay Ready
*2023 – 2025*

- Architected a general-purpose support agent harness steered by composable prompts and skills, enabling rapid rollout of new operational workflows without bespoke pipelines.
- Replaced a failing spam filter with a structured binary LLM classifier in live ticketing, cutting support agent workload by ~60%.
- Shipped a production GPT-4 Vision lease and contract extraction API over 250K+ documents (<1% human correction).
- Built an internal Slack RAG assistant adopted daily across teams in place of a bought tool, cutting manual effort by 60% and retrieval time by 75%.

### Scaled Tooling Analyst — Otter / CloudKitchens ([now Atoms](https://atoms.co))
*2022 – 2023*

- Embedded with operations and engineering to build internal Python data pipelines scoring ~500GB/week.
- Designed and shipped an internal Jira ticketing chatbot from discovery to production, cutting ticket resolution time by 30%.

### Business Analyst — RealPage
*2021 – 2022*

- Built SQL data models and analytics dashboards supporting 75+ internal users, reducing manual reporting effort by 25%.

## Open Source

- **[LiteLLM](https://github.com/BerriAI/litellm/pulls?q=is%3Apr+is%3Amerged+author%3Acolesmcintosh)** (LLM Gateway · Python) — Shipped SSO admin API and UI, Google Cloud Model Armor guardrails, Langfuse OpenTelemetry tracing, usage/health dashboards, provider integrations (Hyperbolic, Lambda AI), Gemini caching and Perplexity search cost accounting, and Ollama streaming fixes.
- **[Pydantic AI](https://github.com/pydantic/pydantic-ai/pulls?q=is%3Apr+is%3Amerged+author%3Acolesmcintosh)** (Agent Framework · Python) — Implemented xAI search tools and attachment lifecycles, model-agnostic X search via subagent fallback, OpenAI input token counting, UI adapter URL sanitization, and provider host/timeout settings.
- **[smolagents](https://github.com/huggingface/smolagents/pulls?q=is%3Apr+is%3Amerged+author%3Acolesmcintosh)** (Hugging Face · Python) — Fixed Enum and custom metaclass execution in sandboxed agent code, added AWS Bedrock API-key support, and refactored MultiStepAgent test coverage.

## Selected Projects

- **[TeraCode](https://github.com/colesmcintosh/teracode)** (Rust · Multi-agent TUI) — Local-first Ratatui switchyard and bounded DAG scheduler over 6 coding-agent CLIs with git worktree isolation, binary-safe patch assembly, SQLite run history, and explicit autonomy gates.
- **[pdfmd](https://github.com/colesmcintosh/pdfmd)** (Rust · Zero-dep Engine) — Zero-dependency PDF-to-Markdown engine built from scratch—PDF parser, DEFLATE decoder, font decoding, and layout reconstruction exposed via CLI and Python bindings.
- **[Openextract](https://github.com/Mellow-Artificial-Intelligence/openextract)** (Python · TypeScript) — Multimodal extraction toolkit converting documents, images, audio, and video into schema-validated Pydantic/Zod objects across 10+ model providers.
- **[Liedar](https://github.com/Mellow-Artificial-Intelligence/liedar)** (TypeScript · Forensics) — Document tampering screening library, CLI, and browser UI combining file metadata, copied-region image detection, and model consistency assessments for human review.
- **[mev](https://huggingface.co/stay-mellow-ai/mev)** (LoRA · Qwen3.5-4B) — Open-weight 4B decision model fine-tuned for 2–24 option structured state selection with calibrated confidence; weights, training notes, and evaluation published on Hugging Face.

## Certifications

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