Open Source

We tested AI agents in four real-world automation scenarios, each requiring different levels of sophistication:

• Customer message routing — classifying support tickets into departments (solvable with keyword matching)
• Multi-turn customer support — handling complete conversations with tool access (requires LLM + tools)
• Contract clause extraction — identifying risk provisions in legal documents (LLMs excel here)
• Server issue remediation — diagnosing and fixing infrastructure problems (dangerous without guardrails)

Each experiment includes evaluation harnesses so you can reproduce results on your own data.

The most comprehensive PDF parser benchmark available. We measured:

• 19 parsers — pymupdf, pdfplumber, marker, docling, Claude, GPT-4o, and more
• 610 documents — contracts, invoices, scientific papers, financial reports, forms, scans
• 6 metrics — text accuracy, structure recovery, table extraction (TEDS), speed

Raw results, evaluation scripts, and full corpus available for independent verification.

One API for 19+ PDF parsing backends. Switch parsers without changing code:

• Smart routing — send scanned docs to OCR backends, tables to pdfplumber, complex layouts to Claude
• Modular install — core has zero heavy dependencies; add backends as needed
• Benchmark-informed — defaults based on PDFbench findings

A plugin system for Claude Code built from 6+ months of daily use. Includes:

• 28 slash commands — structured workflows for explore → plan → implement → review
• 5 specialized agents — architect, test-engineer, code-reviewer, data-scientist, report-generator
• 6 domain skills — writing frameworks, D2 diagrams, ML experimentation

For practitioners who want repeatable agentic workflows, not ad-hoc prompting.

Pre-commit hook that enforces token limits on markdown files. Prevents context window bloat in AI workflows by blocking commits when files exceed thresholds.

Useful for CLAUDE.md files, documentation, and any markdown fed to LLMs.