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 Articles retain their original URLs and publication dates; some historical offers and references are outdated.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
Your LLM-as-a-judge is lying to you (until you calibrate it)
An LLM-as-a-judge is an LLM, so it hallucinates and drifts like the feature it scores. Calibrate it against human labels before you let it gate releases.
Catch LLM regressions in CI before a model update breaks prod
Provider model updates silently shift LLM behavior with no code change on your side. Here is how to build a small golden dataset and a five-minute CI eval that catches th
Why your LLM JSON keeps breaking in production (and the fix)
Unconstrained LLM JSON fails 8 to 15 percent of the time. Here is how schema-constrained decoding cuts parse errors below 0.1 percent, and why valid JSON still is not cor
How to ship an AI feature without wrecking your gross margin
A traditional SaaS product runs at 80%+ gross margin; an AI feature can drag it toward 60%. Protect unit economics in the architecture before you set a price.
Fractional AI engineering team: when it works (and when it doesn't)
A fractional AI team is rarely cheaper per month - it wins on no ramp and lower risk. Here is the decision frame for SaaS teams shipping AI features.
Agentic RAG: when retrieval should be a loop, not a pipeline
Agentic RAG makes retrieval a control loop the model drives, not a single pass. When to use it, how to cap cost and latency, and the stop conditions that keep it bounded.
Semantic caching: cut LLM cost on repeat questions
Semantic caching reuses a stored answer when a new query is close enough in meaning, cutting LLM cost 30-70% on repeat questions. How it differs from prompt caching, and
Self-hosting an LLM vs an API: where the break-even is
The honest break-even math on self-hosting an open LLM versus calling an API: GPU costs, hidden engineering overhead, throughput ceilings, and the token thresholds where
AWS Bedrock vs OpenAI API: the real cost for production
A production-grade cost breakdown of AWS Bedrock vs the OpenAI API: where each wins, the batch and provisioned-throughput levers, and the hidden switching costs.
Designing AI agents that survive prompt injection
Prompt injection cannot be patched away. The fix is architectural: design the agent so an injected instruction has nothing to reach and nowhere to send it.
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.