THE BOUNDEV JOURNAL
The engineering
archive.
Earlier writing on software, hiring, and building AI. Articles retain their original URLs and publication dates; some historical offers and references are outdated.
Browse the archive
LLM provider lock-in: design your product for a model swap
Switching LLM providers should be a config change, not a rewrite. How to design the abstraction seam, what leaks between providers, and when neutrality is worth it.
Load test your AI feature before launch, not after
Load testing an LLM feature is not like a normal API test. Model real prompts, find your 429 ceiling, watch time to first token under load, and turn the numbers into laun
Where the milliseconds go in your RAG pipeline
A stage-by-stage latency budget for a RAG pipeline: how many milliseconds retrieval, reranking, and generation each deserve, and the moves that buy time back to hit your
Right-size LLM reasoning effort to cut cost and latency
Reasoning effort is a per-request dial for LLM cost and latency. Right-size it by task difficulty to cut p95 and spend without dropping quality.
Turn thumbs-down into your AI feature's test suite
A thumbs-down is a labeled production failure. How to route AI feature feedback into an eval set and CI tests that stop the same bug shipping twice.
The audit log your AI feature needs to pass security review
Enterprise deals stall on 'show us what the AI did.' The seven fields an AI feature audit trail needs, and how to build one buyers and the EU AI Act accept.
Your LLM model is being deprecated: the migration plan
OpenAI and Anthropic retire models every quarter, sometimes on two weeks notice. The migration plan we run so a sunset email never breaks a production AI feature.
Why your long-running AI agent loses work when the box restarts
A long agent run dies when a deploy or crash hits mid-flight. Durable execution checkpoints every step so it resumes from where it stopped, exactly once.
Do you need an MCP server, or just function calling?
MCP and function calling are not competitors. A decision guide for when a capability earns its own MCP server and its network hop, and when a plain inline tool is the bet
Your RAG retrieves the wrong chunks? Fix the query first
RAG teams tune chunking, embeddings, and reranking, but send the user's raw question straight to the retriever. Query rewriting, multi-query, and HyDE lift recall without
GO FROM THEORY TO PRACTICE
See an expert review, step by step.
Explore a sample report, a review rubric, and the evidence behind an evaluation.