Software architecture, security, quality engineering, and AI — design notes from production systems
Asset managers are not just fighting yesterday's threats. AI has handed attackers capabilities that scale infinitely. Quantum computing is on the horizon. This is the story of what is actually changing, and what the financial industry needs to do before the window closes.
Ramp data shows U.S. firms paying DeepSeek directly for cheaper inference — but hosted API use routes data to China under PRC law. A technical breakdown of cost pressure, documented risks, self-hosting nuance, and what security claims hold up.
SLOs catch drift in the system you are running. They do not tell you how to deploy schema changes, swap admission rules, or move authorization checks while bookings sit in fan-out pending — without lying to the next customer in line.
Consumer lag can look fine while one booking sits for hours. After you design a Tier-1 write path across multiple services, this is how to measure it in production: what to SLO, what to put on the dashboard, what should page, and when the fix is design instead of runbook.
Two services were each individually correct, and the gap between them became the incident. Action boundaries and idempotent consumers keep single steps honest — not the whole journey. How to draw sync boundaries, model partial failure as state, and make a shared partial-failure vocabulary a Center of Excellence artifact instead of a per-team guess.
Exactly-once delivery is a broker myth, not a system property. Why idempotent consumers, dedupe keys, and fail-closed handling are a design and quality-gate requirement — and why agentic test pipelines inherit the same duplicate-event risk they exist to catch.
A confident model output is not admission to mutate Tier‑1 state. How to design action boundaries when callers are humans or agents — proposal vs admission, risk gates, idempotency, sync vs async, and fail-closed assurance.