Notes from the practical side
of AI and foresight.
Short essays on what actually works when you try to ship machine learning systems.
AI & machine learning
7 postsFrom black box to glass box: why multi-agent systems need to show their work
Three levels of transparency, and what happened when we put each one in front of real users.
→How to actually evaluate an LLM application
Five mistakes I see often in LLM evaluation, and the fixes that have held up in production.
→Why RAG is not a silver bullet
Three failure modes that retrieval does not fix, and what to do about each one.
→Notes from building multi-agent evaluation harnesses
Practical lessons from testing systems where many agents call each other.
→The quiet cost of average accuracy
Why average accuracy hides the failures that matter, and what to track instead.
→Production time series in 2026: the boring tools that win
Unfashionable methods that still run most of what actually matters.
→What air cargo forecasting taught me about real-world ML
Five lessons from a master's thesis, including why your baseline is closer than you think.
→Strategic foresight
6 postsBuilding a trend radar with LLMs (without hallucinating the future)
How JIH Trend Radar uses LLMs only at the final mile, with the heavy lifting done elsewhere.
→What a trend radar actually does in a boardroom
How executive teams really use a trend radar, and what makes one survive the first quarter.
→Bibliometrics for beginners: reading the future in citations
Three things to count, two pitfalls to avoid, and where to start.
→Strategic foresight vs forecasting: they are not the same
Why confusing the two costs strategy teams a quarter, and how to tell them apart.
→The patent signal: what filings quietly reveal
Patents lead products by 2–4 years. Here is how to read them without drowning.
→Designing for the 10-year question
A simple framing for thinking ten years out when you only have data from the last five.
→