Skip to content
Sign in
Artwork for Agentic AI Execution
Agentic AI Execution · June 24 · 30 min

Why Your AI Agent is One API Key Away From a $20,000 Disaster đź’¸

*Are you vibe coding your way into a massive cloud billing nightmare?*AI agents are already writing code, reviewing ads, and making autonomous decisions. But who is actually watching them? In this episode, Leena sits down with Lulu, a data infrastructure expert who has scaled systems at Capital One, Block, and Reddit.They break down the "real talk" of scaling AI safely, why AI-generated code is breaking in production, and how a single unmonitored API key can cause a $20,000 disaster in less than 48 hours. If you are building with LLMs, this breakdown of infrastructure, guardrails, and cost-optimization will save your budget and your production systems.👇 *SUBSCRIBE* for more deep dives into the reality of engineering the AI boom.⏱️ Timestamps & Highlights *00:38* — Lulu’s journey: Inside data teams at Capital One, Block, & Reddit. *01:10* — Why your data infrastructure won't work at AI scale. *02:30* — How Reddit automates 90% of ad reviews using Gen AI enrichment. *04:30* — The hidden cost trap: Snowflake & Databricks vs. LLM scaling. *06:20* — Why running massive frontier models in production will "burn your car". *07:05* — Stop using frontier models for simple tasks (The Small Model Strategy). *09:20* — Why Stanford & Berkeley benchmarks are lying to you about real enterprise costs. *11:15* — What is "Contract-First Development" and why does AI code break? *15:00* — Building "seatbelts" for AI: Secure Kubernetes pods and real-time execution auditing. *18:15* — Startups vs. Enterprises: Compliance, SOC2, and the future of agent security. *21:30* — The Nightmare Scenario: Leaking API keys to your local coding agent. *23:00* — Lightning Round: The one hot take on AI code that engineers HATE. 🔗 🔗 Connections: Connect with Lulu: / luluchengengineeratheart Connect with Leena: / leenaagarwal

0:00-30:27
Play in

transcript

No transcript — this publisher did not publish one.

show notes

*Are you vibe coding your way into a massive cloud billing nightmare?*AI agents are already writing code, reviewing ads, and making autonomous decisions. But who is actually watching them? In this episode, Leena sits down with Lulu, a data infrastructure expert who has scaled systems at Capital One, Block, and Reddit.They break down the "real talk" of scaling AI safely, why AI-generated code is breaking in production, and how a single unmonitored API key can cause a $20,000 disaster in less than 48 hours. If you are building with LLMs, this breakdown of infrastructure, guardrails, and cost-optimization will save your budget and your production systems.👇 *SUBSCRIBE* for more deep dives into the reality of engineering the AI boom.⏱️ Timestamps & Highlights

  • *00:38* — Lulu’s journey: Inside data teams at Capital One, Block, & Reddit.
  • *01:10* — Why your data infrastructure won't work at AI scale.
  • *02:30* — How Reddit automates 90% of ad reviews using Gen AI enrichment.
  • *04:30* — The hidden cost trap: Snowflake & Databricks vs. LLM scaling.
  • *06:20* — Why running massive frontier models in production will "burn your car".
  • *07:05* — Stop using frontier models for simple tasks (The Small Model Strategy).
  • *09:20* — Why Stanford & Berkeley benchmarks are lying to you about real enterprise costs.
  • *11:15* — What is "Contract-First Development" and why does AI code break?
  • *15:00* — Building "seatbelts" for AI: Secure Kubernetes pods and real-time execution auditing.
  • *18:15* — Startups vs. Enterprises: Compliance, SOC2, and the future of agent security.
  • *21:30* — The Nightmare Scenario: Leaking API keys to your local coding agent.
  • *23:00* — Lightning Round: The one hot take on AI code that engineers HATE.

đź”— đź”— Connections:


links14

more episodes

All episodes