OpenTeams

So many people are asking the same question: "I want to run AI locally, what hardware do I actually need?”

Dillon Roach has been having that conversation one-on-one for a while. Now he's put it all in one place.

From the difference between LLM and diffusion model requirements, to why pre-processing time matters for agent and coding workflows, to what the 2026's RAM market means for your build options.

Read it: https://na2.hubs.ly/H068ZHX0

PDF table extraction is hard because the table is not really stored as a table.

The tool has to infer rows, columns, headers, and cell boundaries from text positions on the page. That becomes much harder when the document includes merged cells, nested headers, or packed numeric data.

To compare different approaches, Khuyen Tran tested three Python tools on the same technical PDF:

• Docling, an open-source document converter with a vision-language model pipeline
• Marker, a local PDF-to-Markdown converter built on a multi-stage vision pipeline
• LlamaParse, a cloud-hosted parser that uses an LLM-guided extraction workflow

She evaluated them using two practical criteria: table structure accuracy and processing time.

The article walks through the results with practical examples and diagrams.

🚀 Link: https://na2.hubs.ly/H064NQl0

PDF Table Extraction: Docling vs Marker vs LlamaParse Compared | OpenTeams | AI you own Compare three Python tools for PDF table extraction: Docling, Marker, and LlamaParse. Learn which handles merged cells and multi-level headers best.

Scientific computing and open source innovation are global conversations, and that means they need voices from everywhere.

We supported the Numba Community Workshop in the Democratic Republic of the Congo this May, where local developers, researchers, and students came together to learn, collaborate, and build connections around the Numba ecosystem and open source AI tools.

Big thanks to the organizer, Narcisse Mbunzama, for putting it together. Looking forward to seeing the momentum continue.

At OpenTeams, we believe that the future of open source depends on investing in the next generation of contributors and technology leaders.
This summer, we are excited to welcome 4 interns to our Product Team. Over the coming months, they will work alongside mentors across the entire organization, gaining hands-on experience building open source AI/ML solutions for real-world use cases. We look forward to supporting their professional growth and seeing the impact they make at OpenTeams and beyond!

Because Owned Intelligence isn't just better AI. It's better business.

Find out more at https://na2.hubs.ly/H05YMFY0

Prototype JupyterLab plugins with an AI assistant built in 🧩

Building a JupyterLab plugin can feel overwhelming.

You need to learn the framework, understand the plugin system, and find the right API hook hidden somewhere across the JupyterLab ecosystem.

Plugin Playground now makes that process easier with a built-in AI assistant.

Describe what you want to build, and it can draft the plugin skeleton, connect dependencies, and help you find the right hooks across the JupyterLab API.

With Plugin Playground, you can:
• Turn a plain-English idea into a starting plugin
• Iterate on the code through prompts
• Discover the right JupyterLab APIs faster
• Build in local JupyterLab, Binder, or JupyterLite in the browser

Read the full breakdown by Anuj Singh on the OpenTeams Engineering Blog:

Plugin Playground AI Integration for Faster Plugin Prototyping Learn how we integrated AI into Plugin Playground to help you create, edit, test, package, and share JupyterLab plugins faster in JupyterLite and Binder.

Running local LLM agents safely requires more than giving the model a list of tools.

Even when the model understands the user's intent, it can still pick the wrong file, call the wrong tool, or make a confident mistake during a real workflow.

That is why local LLM agents often need structured tools, scoped permissions, and clear ex*****on boundaries. More of the reliability has to come from the application design, not just the model.

In this article, Nick Byrne explores several design options for that problem:

• Narrow application tools that expose only specific actions
• Workflow graphs that limit which tools are available at each step
• Code mode that lets the model combine approved tools through generated code
• Sandboxing that limits what generated code can read, write, execute, or access

Together, these patterns help make local LLM agents safer and more reliable in real workflows.

🚀 Link: https://na2.hubs.ly/H05Tq6G0

Sandboxing Code Mode for Local LLM Agents Code mode can make local LLM agents more practical, but executing model-written code brings sandboxing back into the architecture.

The promise of AI coding tools is speed.

But faster does not always mean better, especially when teams start skipping the checks that make code reliable.

This article by Johnny Bouder shows how developers can use AI coding tools to move faster without giving up control, quality, or engineering judgment.

🚀 Link: https://na2.hubs.ly/H05NngT0

Slow Down — Simple Lessons for Guiding AI and Shipping Better Code Practical lessons for shipping better code, staying in control, keeping your skills sharp, and getting real value from AI coding tools without losing yourself in the hype.

Happening today at 1 pm ET.

Proving Model Provenance: EO 14365, NIST AI RMF, and Federal Audit Readiness.

Our own Chuck McAndrew + Carahsoft. Part of our ongoing federal series. CPE eligible.

Register and join → https://na2.hubs.ly/H05NpbJ0

A Claude Code skill can turn a simple markdown file into something that feels like working software.

Adam Lewis explored this with a Harvest time-tracking workflow. The skill described the CLI, onboarding flow, billing preferences, and common commands well enough for Claude Code to log, edit, and delete time entries.

That is powerful because there is no packaging, compilation, or deployment step.

But it also exposes the limit of instruction-only systems. A skill can tell an agent what to do, but it cannot enforce what the agent is allowed to access.

This article explains when a skill is enough, when it starts to fail as a security boundary, and how teams can reduce credential exposure with stricter controls.

🚀 Link: https://na2.hubs.ly/H05KVtG0

From Skill to Agent: When a Text File Isn't Enough When does a Claude Code skill stop being enough? See why credential security pushes real workflows toward proper agent architectures.