Builders Unscripted: Ep. 3 - Matias Castello, Product Leader at Alchemy
TL;DR · AI Summary
Alchemy’s product lead Matias Castello shares pivotal AI adoption moments: using Codex in Slack to edit docs directly, and retrospective code review that revealed Codex could have caught a live incident; Datadog data shows >20% of past incidents were preventable—yet this is an incomplete interview transcript with low information density.
Key Takeaways
- Alchemy first integrated Codex into Slack’s docs channel, enabling real-time dev
- A retrospective Codex code review on a past migration-related outage confirmed i
- Datadog reported >20% (1/5) of historical incidents could be auto-captured by Co
Outline
Jump quickly between sections.
Matias Castello, though not trained as an engineer, drives AI tooling adoption at Alchemy through cross-domain experience and iterative practice.
The team added Codex to their internal Slack docs channel, eliminating the need for local site builds and enabling instant documentation tweaks.
After a production incident, the team ran Codex on the faulty code and it correctly identified the race condition—building team confidence.
Datadog reported over 1/5 of past incidents could have been caught by Codex, suggesting current models may prevent 9/10.
Engineers now treat Codex as a virtual reviewer in pull requests, iterating on its suggestions to improve code quality.
Mindmap
See how the topics connect at a glance.
查看大纲文本（无障碍 / 无 JS 友好）
- AI在Alchmy工程实践中的落地路径
- 初期尝试
- Slack + Codex 编辑文档
- 免本地构建，即时修改
- 关键验证
- 回溯式代码审查
- 识别历史竞态条件Bug
- 建立团队信任阈值
- 规模化应用
- PR中嵌入Codex审阅
- 人-AI协同迭代修复
- 行业对标：Datadog 20%+拦截率
Highlights
Key sentences worth saving and sharing.
We added Codex to our Slack docs channel—now we edit docs via AI commands instead of running the full site locally, turning a complex workflow into a one-line request.
After a production incident, the team retroactively ran Codex on the code—and it pinpointed the exact race condition that caused the outage, marking a key trust milestone.
Datadog reported in Jan 2024 that >1/5 of past incidents could have been prevented by Codex; with GPT-4.5, that figure may now reach 9/10.