Posts
Technical writing on software engineering, AI agents, machine learning systems, and engineering leadership. These posts cover practical lessons from building ML platforms, managing engineering teams, and working with cloud infrastructure like AWS and Cloudflare.
Why Bigger Pools Wait LessJun 20, 2026
At the same per-server utilization, a bigger pool of servers queues dramatically less. Drag the sliders to see the economies of scale, decide whether to split or merge your pools, and watch where retries and single-threaded actors quietly steal the win back. Powered by an open-source queueing library.
Testing Whether a Code-Risk Metric Predicts Anything: Defects, Then MaintenanceJun 7, 2026
A risk score is a prediction. I tested riskratchet two ways across 34 libraries: against real bugs, where it came up short, and against future maintenance churn, where it works, carried by a single ingredient. Here is the method, and why shipping zero weight changes was the right call.
Optimizing Our ML Feature Store: Cutting Compute Costs by 55%May 27, 2026
How we migrated from AWS Fargate to Karpenter and used quantile regression to predict memory, reducing compute costs by 55-60% and resource waste by 40% with no application code changes.
Letting Agents Write Code Without Ratcheting Up RiskMay 23, 2026
Why coverage and CRAP are not enough when an agent ships your code, and how I built riskratchet to make function-level risk a per-PR gate.
Five Agents, One Browser: Werewolf on Quack + DuckDBMay 15, 2026
Five LLM agents bluff, scheme, and lynch each other inside your browser tab, each with a private DuckDB-WASM database, federated through a Quack gateway worker. Zero server inference, full row-level auth, and the wolves actually coordinate.
Arrow Flight vs JSON in Next.js: Benchmarking Python and Go Transports for SnowflakeMay 9, 2026
A benchmark of three transport variants between a Snowflake warehouse and a Next.js SSR consumer: JSON over HTTP, Python Arrow Flight, and Go Arrow Flight.
Showing 1–6 of 17 posts