Field notes from production: RAG, MCP, agents, data.
Series · 3 articles
Seven years of startup engineering taught me that the most expensive code is the code that was easy to say yes to. A piece about the price tags nobody puts on the roadmap.
5 min readRunning every LLM call through your best model is the AI equivalent of doing every database query on the primary. Here's how we tier model usage in production, and why it works.
5 min readWhen agents write most of the diff, review stops being a courtesy and becomes the main quality gate. Most teams haven't updated their habits to match.
5 min readEvery team starts by eyeballing outputs. Here's what it takes to graduate from 'looks good to me' to a system you can change without fear.
5 min readThe system prompt stopped being the interesting problem a while ago. What decides whether an agent is useful is everything else you put in the window, and everything you keep out.
5 min readThe most common failure mode in AI projects isn't the model. It's automating a process nobody can actually see.
5 min readHow I run multiple coding agents in parallel without them stepping on each other, and why the bottleneck moved from writing code to reviewing it.
5 min readBuilding an MCP server is a weekend project. Running one in a multi-tenant SaaS, with real users and real ad budgets behind it, is a different job entirely.
5 min readLessons from deploying a RAG pipeline in a regulatory context — from the prototype that impresses to the system that holds under load.
5 min readWhy the ELT approach beats traditional ETL when building product analytics in an early-stage startup — lessons from PulseBoard.
6 min readHow a shared GraphQL schema can serve two products with opposing access patterns — and the pitfalls to avoid when unifying a marketplace and a SaaS tool.
5 min read30 minutes free, no commitment.