Not opinions or transient framework trends. These are the core guidelines I use to design, build, and operate resilient systems.
Every line of code, external dependency, or architectural layer is an operational liability. Simplify systems aggressively. Prefer modular monoliths (moduliths) over distributed microservices until scaling or organizational constraints make a split absolute. Make code deletion a first-class victory.
Leverage SQL, proper indexing, and database-level set operations. Let the database do the filtering, aggregation, and ordering. Network roundtrips, N+1 queries, and in-memory application looping are where backend performance goes to die.
Clean, predictable code that handles failures gracefully is infinitely superior to clever metaprogramming or complex design abstractions. Code is read 10x more than it's written; design for the developer who has to debug it at 3 AM.
A feature is not done until it is observable. If you cannot track its latency, query its execution state, or trace its failure modes in production, you are running blind. Telemetry, structured logs, and tracing are as critical as business logic.
Applied AI is software engineering, not magic. Prompts, LLM execution pipelines, and data embeddings should be version-controlled, schema-validated, cached, and tested in isolation with the exact same rigor applied to database schemas.
Magic (framework defaults, hidden side-effects, implicit global states) hides bugs and compounds technical debt. Build explicit boundaries, establish clear execution flows, and favor predictable systems over clever framework shortcuts.