Braintrust CLI experiments ensure our agents produce the best work.
Eve built Plaintiff Bench: 400+ expert-graded legal tasks.
At Eve, every production mistake becomes a test case.
We get to iterate directly in development and then ship it to production.
Cloudflare scores resolution on every conversation.
Cloudflare's CI/CD evals block regressions before production.
Braintrust is our single source of truth for all of our datasets.
Box gates launches on a 330-question golden dataset.
Box shipped Box Agent once it scored above 80% on core evals.
There are some problems we wouldn't know were problems without Braintrust.
Notion runs regression and frontier evals on every new model.
Notion ships new frontier models to customers in under 24 hours.
Braintrust turns observability into everyday quality in our development process.
Every Pylon prompt needs a playground ID to pass CI.
Automated root cause analysis cut Pylon's on-call debugging to zero.
Braintrust helped us identify several patterns that we wouldn't have found.
Replit streams agent traces into Braintrust in real time.
Replit fixes whole families of problems at once.
We're all proud of what we've put out there and a lot of that is powered by Braintrust.
Bubble tracks quality, latency, and cost on every model swap.
Bubble shipped Agent 2, 1.5x more accurate.
Braintrust is the core of our evaluation framework process.
Navan scores raw call audio with an LLM judge.
Navan evaluates hundreds of daily voice calls above 0.9 macro F1.
We can run hundreds to thousands of experiments with Braintrust.
Dropbox runs a 10,000+ test suite after every merge.
Online scoring catches Dropbox Dash regressions in real time.
We didn't realize we needed deep observability until Braintrust.
Vercel benchmarks every new model before integrating it.
Vercel ships support for new models within hours of release.
Evals give us the confidence to swap models without risking unexpected outcomes.
Rex has traced every agent decision since day one.
Rex ships improvements in 30 minutes.
Loop helps us understand trace details that would be impossible to scan manually.
Retool uses Loop to find error patterns in production logs.
Retool improved accuracy 23% and cut log analysis to minutes.
Braintrust shows us the model's input and output and what actions it took.
Browserbase pairs model and browser observability.
Browserbase's public eval dashboard drew hundreds of thousands of views.