Insights
Writing on data foundations, delivery and what does not work.
- 7 min read
Most digital programmes fail before the first model is trained
The failure is almost never the algorithm. It is that the history the algorithm needed was never recorded.
- 6 min read
On-premise is an architecture, not a compromise
Building for air-gapped deployment from the start produces a different system than porting a cloud product inward.
- 8 min read
What early kick detection actually buys you
Minutes, not miracles — and the size of the influx you have to circulate out is roughly proportional to them.
- 6 min read
Reading forty years of Persian field reports
Persian OCR on clean typed text is solved. The remarks column, handwritten, is not — and that is where the useful detail lives.
- 5 min read
Why ESP run life varies threefold in the same field
Most of the variance is operating practice, and most of the practice is visible in data the operator already has.
- 6 min read
Shadow mode is the only honest way to deploy a model
A model validated on historical data has been validated against the past. Live data disagrees more than anyone expects.