Essays

Streaming Architectures for Teams That Do Not Need Real-Time

Kafka cost $4,200 per month when 94% of consumers queried at hourly or daily granularity. A micro-batch alternative costing $800 per month delivered identical analytical outcomes.

Privacy by Design Is an Architecture Pattern, Not a Checkbox

Systems built with privacy-by-design architectural patterns experience 71% fewer data breach incidents. Privacy is a structural property of architecture, not a feature that can be bolted on.

Time Series Data Requires Its Own Architecture

Migrating time series from PostgreSQL to TimescaleDB reduced query latency by 78% and storage by 62%. Time series access patterns need purpose-built architecture.

The Fairness-Performance Tradeoff Is Real and Underreported

In 3 production fairness projects, I measured accuracy drops of 2.7% to 8.3% when enforcing demographic parity. The tradeoff is real, and honest engagement builds more sustainable fairness than denial.

Open Source AI Ethics: Who Governs Models Without Owners

Open-weight models downloaded over 500 million times have no single entity controlling deployment. When harm occurs, the question of responsibility has no clear answer in current frameworks.

The Ethics of Expertise: When You Know More Than Your Stakeholder

When engineers know more than stakeholders about technical risks, moral obligations arise. 78% of engineering decisions involve information asymmetry. The expert bears the communication burden.

Lakatos and the Research Program of Machine Learning

Lakatos distinguished progressive research programs from degenerative ones. ML has a progressive core of genuine prediction surrounded by a degenerative protective belt of scaling assumptions.

Why smart people build bad systems: The curse of local optimization

Subtraction as Strategy: Lean PMO Design

A three-person PMO outperformed a 30-person one by removing 47 process touchpoints. Subtraction is harder than addition.

Transparency in AI Is a UX Problem, Not Just a Model Problem

Redesigning a SHAP-based explanation interface increased user trust calibration by 52%. AI transparency is an information design problem, not just a model architecture problem.