Latency Labs Blog
Welcome to my blog. I write about low-latency software, quantitative finance, high-performance computing, and AI.
Can an AI Agent Debug a 72% p99 Regression Without Hallucinating?
I built a diagnostic agent with no shell access. Every conclusion must cite machine-generated evidence, and the CI gate stays in deterministic code. Here is what it found, and how I verified it.
EU AI Act Article 50: Transparency as Runtime State
Article 50 became applicable on 2 August 2026. It defines transparency outcomes, not schemas. Here is the provenance and audit architecture I would build to make those outcomes provable.
T+1 Is a Distributed-Systems Problem: Post-Trade Architecture for 2027
ESMA set a 7 December 2026 deadline for allocations and confirmations. Reading the statement as a systems engineer, the interesting problem is not the date field — it is the collapsed recovery window.
Stop Giving AI Agents a Shell: Capability Design for Engineering Tools
A shell collapses hundreds of distinct capabilities into one untyped string. I compare three agent architectures and explain why deterministic tools own the facts.
Your AI Agent Passed the Demo. That Is Not an Evaluation.
Anthropic measured a 6-point benchmark swing from resource configuration alone. Before I build a diagnostic agent, I am designing the evaluation harness that can tell agent failure apart from infrastructure failure.