Insights
Guides written from measured work
No thought leadership. Numbers, methods, and the things that didn't work.
What It Cost to Serve One Million Citation-First RAG Requests on GCP
Measured Cloud Run and Cloud SQL retrieval latency, failures and cost across one million public-edge requests—with the LLM boundary made explicit.
Object Detection Resolution vs Latency: What Our 5,000-Image Test Found
Measured YOLO11 accuracy and p95 latency at 320, 480 and 640 pixels on Apple M4 Pro, plus Cloud Run L4 deployment evidence.
Why Vision Projects Fail on the Floor, Not in the Model
Six ways industrial vision deployments fail after a successful proof of concept — lighting, mounting, class definitions, changeovers, integration and ownership — and how to catch each one during the survey.
Edge or Cloud: How to Decide Where Vision Should Run
A decision framework for where inference belongs — latency budget, connectivity, data-movement restrictions, camera count and maintenance burden — with the questions that settle it.
How Many Images Do You Actually Need to Train a Defect Model?
Why the honest answer is a range, what drives it, and how to build a usable dataset when the process rarely produces the defect you need to catch.
Getting Vision Results Into a PLC Without Upsetting Controls
How to integrate a vision system with plant control equipment: interface design, network segmentation, fail-safe behaviour and the conversations to have with the controls team first.
Noticing a Vision System Has Got Worse, Before Quality Does
Deployed vision systems decay quietly as products, lighting and cameras change. What to monitor, how to sample ground truth affordably, and when retraining is actually warranted.
What Actually Happens During a Vision Site Survey
The two days that determine whether a vision project succeeds: what we measure, what we photograph, who we talk to, and the findings that most often change the plan.
Start with a call, then a costed plan
Thirty minutes on the problem, the site and the constraints. If it looks like a fit, the next step is a four-week AI Pilot at a fixed price, with acceptance criteria signed before any code is written.