Essays
My AI agent tried to keep two EC2 instances alive and terminated them instead
How an SSH quoting mistake made it all the way to EC2 termination, and changed how I think about an agent's blast radius and escape hatch.
Writing
Essays, engineering calls, and a weekly roundup.
What I've learned building data systems and AI agents, the engineering decisions behind Cortrix, and Semantic Tech Insights, my weekly look at semantic technology.
Essays
Semantic Tech Insights · Latest issue
All 3 issuesSemantic Tech Insights
Agent Memory Starts Dreaming in the Background: When to Consolidate, Who Signs Off, and Whether the Originals Stay
Databricks consolidates agent memory in the background, PowerContext waits for a person to approve each revision, and EverOS treats the original Markdown as the source of truth. Two papers split on whether to keep the raw records.
Cortrix in practice
Cortrix in practice
Agent First Design: kill -9 Halfway Through a Write, and the Data Lines Up Again After Restart
We killed Cortrix with kill -9 halfway through importing 40 documents. After the restart, every vector matched a block again. Here is what it checks and how to run the test yourself.
Cortrix in practice
Stop Making Claude Code Carry Your Whole Project: Up to 43% Fewer Tokens with Cortrix
In our test on 932 internal design docs, Claude Code scored the same with and without Cortrix, but used 43% fewer tokens on the hardest question and 32% fewer tool calls.
Cortrix in practice
Agent First Design: HTTP 200 Doesn't Mean You Got Everything
When a Cortrix query can't reach part of what it asked for, it still returns HTTP 200 and says what's missing. See what comes back, how an agent should use it, and when to treat it as a failure.
Cortrix in practice
Write-Time LLM Inference Gives Every Query More Recall Paths
Cortrix runs an LLM when a document is written, so context, likely questions, and a document summary each become a recall path at query time. See what we measured and when to enable it.
Cortrix in practice
Cortrix Handles Retrieval. Jev or Laya Can Help the Agent Decide.
Cortrix retrieves relevant material and preserves its sources. An agent combines that with current business state, then can use Jev or Laya to assess its next step.
Cortrix in practice
Jev and Laya are getting hyped. We're passing for now.
Jev can judge a candidate set, but it cannot improve recall. I looked at the public evidence and decided not to replace Cortrix's local BGE path.
Cortrix in practice
Vector is not semantic
A small retrieval demo separates similarity, ranking, lifecycle checks, and the application rule that decides whether a result is usable.
Paper notes
DBpedia explained: why linked data still matters
Why stable identities and linked data still matter when your agents retrieve knowledge from different systems.
Paper notes
TransE explained: relationships beyond vector similarity
What the simple h + r ≈ t formula tells us about relationships in a knowledge graph, and what a similarity score leaves out.
Paper notes
RAG explained: external memory brings data governance back
RAG lets a model use external knowledge. The original paper also helps explain why sources, updates, and data governance still matter.
Paper notes
Reflexion explained: agent memory without weight updates
Reflexion turns feedback into memory without changing model weights. Its most useful lesson may be about the evaluator.
Paper notes
HippoRAG explained: why RAG comes back to knowledge graphs
HippoRAG connects clues across documents with a knowledge graph. Here is how it works, and where the graph can go wrong.
Cortrix in practice
Cortrix RC2: 24.4% higher FiQA recall, with Apache 2.0
Cortrix RC2 brings 24.4% higher FiQA Recall@10 in the published comparison, Apache 2.0 licensing, and practical ingestion and agent integration fixes.