Slotstream
Open sourceRun a 105 GB AI model on a Mac that can't hold it. A native Swift engine.

Software & AI engineer and builder · Colombia
I build AI products and the systems behind them.
I've shipped software since 2010 and worked deeply in AI since early 2022, from production agents to model training and inference. Right now, I'm building Sevra's local AI workbook and maintaining the open-source Slotstream engine.
Run a 105 GB AI model on a Mac that can't hold it. A native Swift engine.
I built a chat interface over three trucking-industry data products for Randall Reilly.
A database made of plain files, built for agents to read and maintain. An open format with a Rust toolkit.
Talk to a language model. Teach it something new. Take it apart and see what changes. Runs in your browser.
An Android-based phone OS configured through a Markdown file. A working prototype on a Nothing Phone 3.
An open-source Mac workbook for documents, tables and local AI. Available as a source preview.
“Stop considering and work with Carlos. He’s been an incredible asset to my business, clear on strategy, excellent communicator, and the rare kind of builder who can actually ship. He led, contributed, and taught along the way, and brought real passion to every project.”
Will McLellarn
Founder, General Dev · Stanford alum
Our work on Stratas Assistant
“Highly recommend Carlos – he helped unblock me and gave me a path forward in my MVP with Bolt, and then he even provided suggestions for ways to streamline the process as I moved forward.”
Naveed Hedayati
Head of Product, Co-Intelligence Labs
My recent client work includes AI assistants at Katapult Labs, model routing and backend development at Evermuse, and Stratas Assistant for Randall Reilly. I've led teams of 4–12.
I built a government-contract discovery SaaS with Go, React and Postgres, then helped move a video-communications product from Firebase to a production Node.js/Postgres backend at Tribe. My open-source work goes back to Blockly's zoom controls in 2015.
Engineering experienceA complete computer for AI agents: dedicated machines where agents used real browsers, terminals, and filesystems.
This was the earlier product. The name now belongs to a separate service, VibeCraft, my AI Operations Partner.
Encrypt two numbers in your browser, compute on their ciphertexts in a separate worker, then decrypt the answer. Real fully homomorphic encryption with live metrics and the client/server roles made visible.
An interactive walkthrough of GPT-2 from first principles. Built so I could finally understand transformers properly.
A visual, interactive walkthrough of the core idea behind GraphRAG and why it works differently.
Claude Computer Use as a desktop app. Built with Tauri and the Vercel AI SDK.
An AI therapist with a whispering voice. A demo using the OpenAI Realtime API, ElevenLabs, and Fal.ai.
How to build a JavaScript runtime on top of Deno Core and Rust.
Running multiple Deno tasks safely inside a Tauri app. Permissions, IO, and lifecycle.
An earlier platform for building apps with natural language. Predates a lot of what's happening now.
Unofficial Figma desktop app for Linux, with local font support. Built it because I needed it.
JavaScript to Blockly blocks. Parses code and imports it into the visual block editor.
I help teams build AI products and fix failures in live agent workflows. We pick a measurable problem, implement the change and check whether it worked. For ongoing AI operations, my current service is VibeCraft.