David Kim

Design engineer @braintrustdata

[David Kim @dvddkkim]

@parkerhendo senpai

[Anna @annawangg]

NYC, we are inside you!
Surreal to see @braintrust x @mta in the wild right after moving here. If you’re in New York, keep an eye out
Also, we're hiring a bunch for our NY + SF offices :)

[David Kim @dvddkkim]

@sharanjhangiani @braintrust @ankrgyl Sharan!

[David Kim @dvddkkim]

@parkerhendo For me it’s when I sit crunched like a goblin w my laptop on my knees

[David Kim @dvddkkim]

I compared Kimi K2.5 to Sonnet 4.5 to see which is better at prompt optimization using @braintrust

I gave it 4 scenarios -- ranging from movie guessing to sql gen.

Sonnet came out slightly on top, improving score by 54.6% on average across multiple runs.

However, time to first token for Kimi K2.5 was 3 times faster than Sonnet

[David Kim @dvddkkim]

We recently added support for SQL at @braintrust !

I thought our agent, Loop, would be better out of the box at generating SQL over our own query language (BTQL), given that's what models are trained on. Surprisingly, this wasn't the case because our harness was optimized for generating BTQL.

With some tweaks to our tool implementation and prompt, guided by evals, we were able to turn that around!

[David Kim @dvddkkim]

Seeing actionable error messages help with steering the model so much.

error message like
"error: parse error"

doesn't tell the agent what they need to do next but "error:parse error. Use X instead" does wonders

[Braintrust @braintrust]

Brainstore is the purpose-built database for AI observability.

It handles the volume and complexity of modern AI systems, so you can run eval workflows at scale.

And it makes Braintrust 24x faster than competitors.

[morgane palomares @morgane_paloma]

Come to Trace, @braintrust's first user conference:
* Feb 25, 2026
* California Academy of Sciences in SF
* 500 AI builders
* Eval workshops, demos, sessions from AI leaders

No boring talks. Bring your laptop → https://t.co/eFkjC0Zh0l :)

[David Kim @dvddkkim]

gpt-5.2 performed slightly better on our prompt optimization evals but gpt-5.1 was faster and more token efficient

evals powered by @braintrust

[David Kim @dvddkkim]

@realmcore_ @braintrust what have you had problems with? happy to help!

[David Kim @dvddkkim]

@realmcore_ @braintrust I've been dogfooding Braintrust to eval our agent and it's been working well for me!

For our agent, i do most of the eval through the SDK. We built a cli version of our agent, set up an eval env using the sdk and throw whatever eval task at it

[David Kim @dvddkkim]

It was so fun working on this with the team :)

If you want to work on AI tooling and/or AI where you have autonomy with teammates who trust and support you, we are hiring for all roles!

(If you are a designer who codes interested -- I'm happy to chat and share my exp. dm me!)

[Braintrust @braintrust]

Production AI breaks in unexpected ways.

Loop helps you figure out what happened and what to fix, fast.

[Braintrust @braintrust]

Production AI breaks in unexpected ways.

Loop helps you figure out what happened and what to fix, fast.

[David Kim @dvddkkim]

So much UX possibilities with low latency models

[David Kim @dvddkkim]

@parkerhendo @braintrustdata So happy to be working with you :) !!

[David Kim @dvddkkim]

@reneeshah123 @braintrustdata https://t.co/yYFKhdWFKP

[David Kim @dvddkkim]

Thought I borked the system prompt but it was just anthropic going down

[David Kim @dvddkkim]

o3 is the bro I'm studying for my cs exam with. Claude 4 is the bro I'm pulling up to hoop with