I've spent 12+ years in SRE. The best fix I ever shipped wasn't clever. It took a dev environment setup from 3 days down to under 15 minutes.
Now I'm building tools that do that for CI: HeyEddi CI cut our CI compute by 70%.
I'll share what I've learned here, building in public from México 🇲🇽
Devs en LATAM: dejen de venderse como "recursos".
Lo que más me ha abierto puertas no es otro certificado. Es poder decir qué cambió gracias a mi trabajo: "el setup pasó de días a minutos" pesa más que "sé Terraform".
Midan lo que hacen. Cuéntenlo. 🇲🇽
1. Hot take: most "AI reliability" problems are just old ops problems wearing a new outfit 🧵
2. Cost is part of uptime. An AI feature that works but burns 10x its budget is getting turned off by finance next month. Put the spending cap in from day one.
3. Trust the evidence, not the confidence. If a model can't point to where its answer came from, treat it as a guess.
4. Keep a human on the irreversible stuff: deploys, merges, deletes. Automate everything up to the button, and leave the button to a person.
5. Boring wins. Infra in code, throwaway workers, clear owners. The best fix I ever shipped was the least clever one.
6. Don't marry a vendor. Clouds taught us that, and models are teaching it again.
7. That's how I'm building HeyEddi CI. Curious what you'd add 👇
https://t.co/l9hRK9vvoD
@Surendar__05 Dumb people make dumb choices.
Seriously, the PC begind macOs is not bad, was never bad.
It wasn't even expensive... On certain parts of the world.
@ayesha_fatiima Gitlab? And a bunch of opensource ones too.
I have been using GitHub due to some integrations, a PR reviewer. It died, and I built my own on https://t.co/rv5iJGXnBw, I might go back to gitlab, now that you made me think about it.
@ulrick@SebastianHold Un agente que comenta en cada PR y se equivoca la mitad de las veces te cuesta más que no tenerlo. Nosotros no publicamos nada que no pueda citar archivo y línea.
@brianrequarth@velez_david Nubank made 'built in LatAm, sold everywhere' believable. A lot of us building devtools from Mexico are betting on the same path.
@garrytan Saw the same thing in CI. Plain rules handle most failures before a model touches them, and the model only gets the leftovers. Cheaper and way more repeatable.