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 Articles retain their original URLs and publication dates; some historical offers and references are outdated.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
When to hire your first in-house AI engineer (and when not to)
The signals that justify a full-time AI engineer over a fractional or subscription team, the cost math that has to clear first, and how to time the switch without a deliv
The AI gateway: one path for every LLM call in production
Every LLM call in production should route through one AI gateway that owns rate limits, retries, caching, routing, and cost. Here is why, and what it must do.
When to replace a frontier API call with a small fine-tuned model
A frontier API is the right default, but a few high-volume, narrow calls in your product are worth graduating to a small fine-tuned model you own. Here is which ones qual
Your LLM bill is mostly output tokens, not input
Trimming the prompt is the wrong end. Output tokens cost several times more than input, so controlling response length is usually the fastest LLM cost lever you have.
Self-hosting an LLM is cheaper than the API far later than you think
Open models now match the frontier, so self-hosting is a pure cost question. The GPU price is the smallest number, and break-even lands far later than you think.
Most LLM queries do not need your most expensive model
One frontier model on every request means paying 15 to 20x more than most queries need. Here is how model routing and cascades cut LLM spend 45 to 85 percent.
The guardrails an AI feature needs before it touches a customer
A feature that behaves in a demo is not ready for strangers. The input, output, and containment guardrails, plus the cost cap, timeout, and fallback most teams forget.
Run a 2-day feasibility spike before you promise an AI feature
A two-day feasibility spike tells you whether an AI feature is shippable on your real data before it eats a sprint. Real examples, a plain prompt, and a go or no-go pass
Stop hardcoding prompts: version them like production code
Hardcoded prompts ship silent regressions. Treat each prompt as a versioned, eval-gated artifact, stamp every output with its version, and roll back without a code deploy
Your MCP server passes manual tests and fails in agents
Five tool calls at 71 percent each succeed end-to-end only 18 percent of the time. The MCP server failure modes that break agents, and how to fix each.
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.