Ankur Goyal

Come learn how we eval our thing that evals your things!

things i install when i ssh: * tmux * neovim * claude * uv * gh * htop * dstat

few things are more satisfying in systems than getting rid of a dependency

CI environments are about to be the new hot thing Who is going to make it super easy to quickly give your agent a sandbox with your whole dev setup so it can iterate on its own, outside of your computer? Not an easy problem to nail...

Let AI generate the code or let AI generate the code that generates the code

people seem to really love Loop. every day, people ask, "how do I build an agent like this?" the answer is simple :) use @braintrustdata

I often talk to people ask "so Braintrust... it's an eval framework right?" The answer is no :). Braintrust is infrastructure for building, scaling, and optimizing AI evaluations FWIW eval frameworks are great. Most of them support braintrust! braintrust.dev/blog/braintrus…

Now with a notebook you can follow to run the experiments in your own @braintrustdata org ….

There's a very important mental shift when you transition from "AI can't do X for me" to "I'm too dumb to figure out how to do AI for X".

At present, MCP is a plugin system for Cursor. I'm interested to see if it finds a life beyond that. Or we just write better APIs/CLIs that are easier for LLMs to understand.

My current framework for leveraging AI is combatting laziness. If I think of something that seems important, but I feel too lazy to do it, I ask "why am I not solving this with AI?"

Gentle reminder We're hiring across a lot of roles, including systems (brainstore), product/design eng, infra (help us make aws/azure/gcp/k8s deployments gr8), support, growth, sales, BD, SE you name it. Anyone who wants to chat about roles this weekend, I'm around

If you spend 30 seconds thinking about it... no shit LLMs are good at SEO...

Sometimes I hand code things specifically so that I understand the implementation deeply. This is mostly so that I can more effectively maintain it in the future. It's interesting how the calculus might evolve over time for when you should opt-into hand coding things.

i would suggest having two programming environments: sync and async. The sync one should use an IDE of your choice and optimize for your synchronous attention. The async one should use a background agent of your choice and let you work on a loop of request, check in a few mins…

I think one of the most interesting opportunities over the next decade will be building simple, elegant systems that LLMs can use and abstract away hard computer science problems. I've noticed over the past several months that tool use is prompt engineering. You HAVE to create…

o3 formats SQL queries extremely pleasurably

the hardest problem in o11y is "why didn't my ___ show up?"

the otel protobufs are actually a lot nicer than the otel libraries

I hate the fact that today the choice is between Great AI and Great Text Editor. VSCode seems so ripe for disruption.