Writing
Notes on engineering leadership, agentic AI, and building platforms at scale.
- 2026.09.19 The Harness Decides When the Work Is Done
What harness optimization research teaches us about completion checks, verified repository memory, and improving coding agents through controlled experiments.
- 2026.09.11 Serving LLMs at Scale with vLLM and SGLang
A practical guide to inference architecture, KV cache economics, prefix reuse, routing, and operating vLLM and SGLang under production latency targets.
- 2026.09.05 From Copilot to Delegate: Why Autonomous Agents Need a Sandbox
Reliable AI delegation needs more than a capable model. It needs an isolated execution boundary for context, access, workflow, cost, and human escalation.
- 2026.08.28 A Trillion Databases, and Not One for Analytics: Why DuckDB Exists
In 2019, SQLite had over a trillion deployments and analytics had nothing like it. A SIGMOD paper named the gap, DuckDB filled it - and hardware quietly made single-machine analytics the default.
- 2026.08.22 0.5 MB per Token: How the KV Cache Makes LLM Inference Possible
Every token an LLM generates must attend over everything before it. The KV cache is why that doesn't collapse under its own cost - and why GPU memory, not compute, is the real bottleneck in inference.
- 2026.08.19 2.3 Million Queries a Second on Plain Old MySQL: How Slack Does It
Slack runs its message store on MySQL at 2ms median latency - because Vitess makes sharding someone else's problem. What that teaches us about scaling databases.
- 2026.08.18 Message Queues: The Shock Absorbers of Distributed Systems
When to reach for a queue, how the ack loop actually works, and how partitioning scales consumption - without creating hot partitions.