In second part of July I explored the practical side of building systems that stay correct over time. A common thread was the gap between an AI or backend system working in development and staying reliable under real load.
Several articles focused on keeping AI development disciplined: maintaining compiled knowledge, keeping specs and tests in sync, and choosing the right LLM serving stack. Another thread was classic distributed systems reliability, with a deep dive into dead-letter queues.
1 Keeping AI knowledge and development disciplined
Two articles look at the engineering practices that keep AI systems correct as they evolve. The focus is on detection, traceability, and CI enforcement rather than manual review.
- LLM Wiki Maintenance: Drift, Contradictions and Review
- Keeping Specs, Tests, And Code In Sync In AI Development
2 Moving from Ollama to vLLM
Not every local LLM server needs to stay on Ollama forever. This article covers when the migration signals appear, how to plan the move, and a practical Docker Compose setup for vLLM.
3 Handling poison messages with dead-letter queues
Dead-letter queues are one of those patterns that look simple until you need to design replay, retention, and alerting around them. This article covers the tradeoffs between retry and discard policies, plus safe replay strategies.
If one of these articles is useful to someone building AI systems, backend systems, infrastructure, or technical knowledge workflows, please forward this email or share the link with them.
Thanks for reading, Rost