· 18 min read
Building Chat Memory That Still Works at Turn 900
Replaying the full chat history works at turn 3. At turn 900 it is why your bill is high and your assistant is wrong.
Read on Substack →
ai engineeringsystem design
Featured writing
Writing on AI engineering
Prompt engineering, system design, and building AI products that hold up past the demo. Cross-posted from Substack.
· 6 min read
The Math Behind Your Feed’s Order (And What It Says About the Product)
Every social feed you scroll runs a scoring function on every post, in real time, before you ever see it.
Read on Substack →
ai engineeringsystem design
· 2 min read
AI Engineering Is Real Work. That’s Exactly Why It Won’t Stay a Category.
Last weekend I spoke at CodeCon Summit. The talk was called “O Guia do Mochileiro da Engenharia de IA” — the hitchhiker’s guide to AI engineering — and the premise was ambitious: walk through 42 concepts anyone entering this field needs to understand today, in one session, for a room of over 300 engineers.
Read on Substack →
ai engineeringsoftware engineering
· 5 min read
MCP + Skills: The New Interface for Engineering Work
There’s a type of cognitive overhead that engineers have normalized so completely they stopped noticing it.
Read on Substack →
ai engineeringharness engineering
· 4 min read
Streaming in LLM Agents Isn’t Just a UX Feature. It’s an Architecture Decision.
Most teams add streaming to their AI products because it feels faster.
Read on Substack →
ai engineeringsystem design
· 7 min read
Building Internal Chatbots That Actually Work
Every company I’ve worked with has the same problem dressed in different clothes.
Read on Substack →
ragai engineeringsystem design
· 12 min read
The Unit Economics of AI Products
Traditional SaaS has a comforting property: infrastructure costs are largely fixed.
Read on Substack →
ai engineering
· 18 min read
Caveman Prompting: I Tested the Claim. The Real Finding Is Somewhere Else.
There’s an idea that circulates in AI engineering circles called “caveman prompting.” The premise: write your instructions in stripped-down, telegraphic English — almost no grammar, no courtesy, just the bare intent — and the model will still deliver.
Read on Substack →
prompt engineeringai engineering