← All speakers

Hassan El Mghari

Conference affiliation: Director of Developer Experience · Together AI · 2026

Hassan El Mghari leads developer experience at Together AI and builds open-source applications that people can use without knowing which model runs them. His RoomGPT grew out of a specific frustration with AI room redesigns. An image generator could produce an attractive living room while changing its depth and dimensions. Hassan chose ControlNet so someone could try different furniture, tiles and colors and still recognize their own room. By October 2023, more than two million people had registered and used it. The original implementation remains open source, and RoomGPT.io developed into a paid service.

On this page

The weekend projects at Vercel

Hassan began learning to code in 2020, switched his college concentration from electrical to computer engineering and graduated in 2021. After an early developer advocacy role, he joined Vercel that September and became a senior developer advocate the following year. His first AI project there solved a publishing problem. His team had several hundred conference photos that needed alt text. He used an image-description service, reviewed its suggestions, corrected a couple of descriptions and released the code. The Weekend AI Engineer

Hassan often built the smallest version of an idea he could try over a weekend. When a friend wanted help writing Git commit messages, Hassan wrote a script in under two hours. Once other developers showed interest, he packaged it as AI Commits, a command-line tool that reads staged changes and proposes a message for the developer to review. Riccardo Giorato and Hiroki Osame maintain it alongside him.

RestorePhotos began with old family photographs, including one of his father doing karate with his face blurred. A single call to GFPGAN, a face-restoration model, was enough to sharpen the face. Running the app for the public took more work. When people began uploading inappropriate images, Hassan added a screening step before restoration and published the screening code as a library.

Teaching through working applications

After a period consulting on AI applications, Hassan joined Together AI in January 2024 to lead developer relations. His role grew to include a developer experience team working on documentation, demos and developer education. One of his projects there, LlamaCoder, used Meta’s Llama 3.1 405B with Together AI’s hosted inference. A user could describe a small app and try it in the browser. Hassan split the work into two model calls, one to plan the app and another to turn the plan into React code. In 2025, he estimated that users had generated about a million apps. With BlinkShot, he chose Flux Schnell for a different reason: it was fast enough to generate images as the user typed. Using OSS models to build AI apps with millions of users

Hassan favors launching with an existing model while the cost of testing an idea is still a weekend’s work. He returns to apps when a better model makes a feature possible or improves the result. Publishing the implementation serves his teaching as well as his distribution: developers can learn from the code, suggest changes and share a working app. To keep projects free to try, he sets usage limits according to the cost per request and arranges sponsorships. Together AI has supplied inference, Neon database service and Clerk authentication. Each provider gets a public example that developers can run themselves.

Designing the interface after the first draft

Mobile traffic surprised Hassan when he began looking at how people used his apps. He estimates that he spends about 80 percent of his time on the interface for most of them. He collects screenshots of designs he likes and gives coding agents examples to work from. For subsequent edits, he often uses a faster, cheaper open model that has done well on the particular design task. That makes it practical to try several versions of a detail such as the loading state someone sees while an image is being generated. The Missing Layer: Design Taste in AI Agents

Hassan built Hallmark at Together AI to make those design choices reusable. It pairs rules about recurring habits, such as widely spaced uppercase labels above every heading, with visual themes that show an agent what to aim for. Its browser checks distinguish a broken phone layout, which needs fixing, from a stylistic default a designer can choose deliberately. Hassan keeps his own recurring preferences in AGENTS.md. His open-source /variate skill helps when he is still choosing a direction: it generates alternatives for the same interface element, such as a navigation bar, so a developer can compare them in context.

A forecast his father could audit

Hassan built a Costco forecasting app with his father, Omar El Mghari, a former CFO and head of internal audit. Hassan had the model produce formulas the app could run and displayed them in a spreadsheet for review. The forecast looked close to Costco’s reported results, but Omar found that errors in the model partly canceled each other out. They turned some of Omar’s accounting checks into tests the app could run on later forecasts.

Read the topics behind these talks

4 conference talks

Loading more talks and highlights…

References