David Sancho

David Sancho

@david.sancho.dev

Open Source UI infra at @ahrefs.bsky.social with OCaml Made styled-ppx and server-reason-react Working on Reason / Melange / reason-react and friends Made styled-ppx and server-reason-react Working on reason-react / Melange / Reason

2K followers1K following1.1K posts

Pinned

I'm david. Software dev. Can't shut up. I have problems writing in English and didn't get a proper education. I speak to fast, in spanish probably faster. Love OCaml, prefer Reason. enjoy React, CSS, and the frontend part of my job. Wrote a CSS parser and a CSS type-checker.

I was interviewing someone with 4 years of OCaml experience, and said: "The language solved problems I didn't know I had, but the tooling invented problems I didn't know were possible"

Previously, it was part of server-reason-react. Now that the single context from dune is becoming real, it feels the right time to formalize the share code story. ppx_platform is like ppx_optcomp but for Melange/Native

Martin Odersky on safer agents: don't make the agent codemode directly. Compile it and enforce it with a capability library, then run it. The type checker decides what the agent is allowed to do. I imagine all agents will implement this sooner o later

the harnesses (dot, grok bot, etc) will become the OS. Both a deterministic launcher to run apps, but as well interfaces for llms with toolcals, mcps and all our life

there will be a difference between a live app and a tool app. A live app, is the one that you will open, the other will be used by the llm for you

harnesses should enable a hook to run auto-improve that suggests improvements in your program or in your environment to help you get better results: - propose tests, skills, different models for each task, etc - learns how you fail at some tasks, and when you succeed

- removes the need for openai or claude to "please update your skills, our model is much better, balbla", it will be done automatically auto-improve is currently a skill I made for myself that runs on the background and spawns reports

CSS at 0 context window > display: flex; > align-items: center; > justify-content: center; CSS after compaction > position:absolute; > left: -18.245555px; > top: -24.99999998px;; > margin-left:5.333333px; /* ajdustment */ > margin-top:…

me: I work on developer tooling wise person: wait, do you mean the migrations team? me: yeah 😢 wise person: ah, so you break everyone's build on a schedule me: yeah 😢 wise person: so nobody notices your work until it's broken me: yeah 😢 wise person: looks fun me: fml 😭

Mandatory reminder to add the :gh-down: as an emoji to Slack so it's easier to communicate when all links to issues and prs are broken and you are mad

hey gpt, can you review the changes made by the thermo-nuclear-review, from the comments by Rabbit AI in the pr of the review of the audit of the architecture

I need a hater, a friendly hater, but a hater. Someone who I make a post or a tweet and he dms me: "this is shit" and runs away after 2 sentences. but not a generic hater, someone who understands my shit and can argue correctly from another pov

over the moon rn, about to land one of the biggest migrations I've ever done... a PR that's been open since 11th of November ripped the emotion runtime out of styled-ppx and moved into a static extraction at compile-time. super fun to build. but the migration was a big mountain

David Sancho reposted

Hey! I am open for contract or full-time work as a full-stack developer. I've worked on industrial software, real estate tech, and a visual website builder. I specialize in design systems, a11y, and UX. Previously maintained PostCSS, Rollup, and SVGO. Currently building weareonhire! on Open Social

After the AI Engineer conference, I'm thinking about some problems from AI models to overcome. I got this list from the top of my head. Anything missing? • continuous learning or better memory? • long-term planning • prompt quality resilience: different prompts shouldn’t alter the outcome cont

• pushback: they shouldn’t agree in all cases and infer the real goal • multi-step reliability • cost aware: they should know when writing a Python script is better than doing it themselves