Notes on applied AI, software engineering, and technology.
Keep serverless costs predictable with concurrency limits, budgets, tags, and cost-per-request tracking.
16 min read
Use multiple agents only when one cannot do the job, with clear ownership, budgets, and stop conditions.
18 min read
Tune HNSW, filters, and vector precision against recall and latency, not defaults.
16 min read
Build an idempotent C# Service Bus worker with sessions, explicit settlement, and dead-letter recovery.
18 min read
Make only the right parts portable, and decide where your data lives first.
14 min read
Validate every agent tool call before execution, then make retries bounded and side effects idempotent.
21 min read
Preserve document structure when chunking so retrieval keeps the context tables and policies need.
19 min read
Bound retries with timeouts and circuit breakers so a failing dependency cannot take down your API.
18 min read
Choose Service Bus, Event Grid, or Event Hubs by whether you send commands, events, or streams.
16 min read
Build a Python MCP server that exposes tools and data to compatible AI clients.
17 min read