Writing
Notes from systems meeting real devices.
Notes on payment systems, device software, transaction uncertainty, platform design, and the emerging role of AI in commerce. I write from the questions that appear in real operating conditions.
Consent, mandates, and trust in autonomous agent payments
How an agent can pay on your behalf and audibly prove why. A short taxonomy of the guardrails, and where each is easy to get wrong.
Agent payment protocols: what's converging
Multiple industry proposals for how AI agents should initiate payments have emerged in 2025-2026. Where they agree and where they don't.
Grounding LLM answers with source citations
For commerce applications, an LLM answer without a citation is not an answer — it's a liability. The engineering pattern that actually works.
Model budget controls in customer-facing AI
Every LLM feature has an unbounded cost tail. The controls that prevent a support agent from spending your quarter's budget in an afternoon.
AI-drafted reconciliation reports
LLMs are good at turning structured data into readable prose. Reconciliation summary reporting is a well-shaped use case, with narrow guardrails.
Explaining a declined transaction with an LLM
A merchant asking "why was this declined?" is a common support ticket. What a well-designed LLM assist looks like — and where it must refuse.
RAG over compliance documents
Compliance documents change slowly, are consulted often, and require precise citation. A well-shaped RAG use case with specific engineering demands.
Tool-calling for merchant support agents
Merchant support workflows involve looking things up, running checks, and drafting responses. Tool-calling turns an LLM into a first-line agent for well-shaped tasks.
The prompt injection surface in commerce
Anywhere a merchant or customer supplies input that reaches an LLM, prompt injection is a live threat. A field taxonomy of the attack surface.
Why AI won't replace fraud analysts (yet)
Fraud detection is one of the areas where AI hype is loudest. What actually improves fraud analyst productivity, and what the models still can't do.