Eric Le I build open-source tools for coding with AI agents
I'm Eric, a software developer in Gothenburg. I studied engineering physics at Chalmers and built my first agent during my master's thesis at Volvo. These days I build open-source tools for coding with AI agents, because an agent does its best work when the plan, the rules and the review around it are set up well, and that setup is the part I enjoy.
Outside of that I play tennis and travel whenever I can.
Where I've worked and studied
dec 2025 to present
sep to dec 2025
learning web development
Worked through most of The Odin Project’s full-stack JavaScript path, from HTML and CSS to React, then started Stackr in TypeScript that December.
jan 2024 to jun 2025
volvo technology, gothenburg
Built an agent that let test engineers get plots and figures from their truck test data by asking in plain language. I built and deployed it, starting from my master’s thesis. Presented it remotely at Generative AI for Automotive USA 2024, an industry conference.
jun to aug 2023
internship at bac ha software, hanoi
Reproduced published GAN architectures in TensorFlow for image generation and image-to-image translation, and trained them from scratch on the company’s GPU server.
sep 2019 to jun 2024
msc in engineering, engineering physics, chalmers
Engineering physics first, then a master’s in complex adaptive systems, where simple agents and their environment produce behavior nobody designed.
I wanted every project to work the same way with AI coding agents. canon takes each change from plan to reviewed pull request, with the same skills and rules in every repository. I use it every day. On npm as @erclx/canon, with a live build at canon.erclx.dev.
I wanted my agents to design in something I could see and change. Frameboard is a board where they draw interfaces as plain HTML files, and I steer them by hand or with notes they answer. I use it daily for sketches, design systems and UI fixes. On npm, with a page at frameboard.erclx.dev.
Staging files by hand for every prompt was the slowest part of every planning chat. VS Code extension that stages files across a workspace into one block of LLM context. Installed from the VS Code Marketplace and Open VSX.
I kept retyping the same prompts across three different chat windows. Chrome extension that saves prompts and drops them into Claude, Gemini, and ChatGPT from the chat input. Live on the Chrome Web Store.
I wanted to know whether retrieval still helped when a model could read the whole EU AI Act. An agent that maps a described system to the articles that apply, cited to the text. Reading the whole Act found every provision. Search plus cross-reference traversal found 83% of them for about a seventh of the tokens, and showed which text it used.
Open to work
Open to developer roles in Gothenburg or remote, full-time or contract.
