Human Review Is a Feature, Not a Failure.
Confidence, explainability, and traceability are most useful when AI supports informed human judgment rather than replacing it.
Lab Notes
Technical notes from real builds.
Focused observations, lessons, and technical memos from applied AI projects. These notes are not blog posts. They are practical records of what matters when AI systems meet real-world constraints.
Products Tell Stories. Projects Tell Histories.
Why product pages should explain present value while engineering records preserve the path that produced it.
Evidence Before Intelligence
Trustworthy manufacturing AI begins by separating measured facts, analytical observations, and interpretations.
Privacy-Safe Demos Without Fake Engineering
Public demonstrations should protect production data while preserving the relationships that make the underlying engineering meaningful.
From Dashboard to Investigation Workspace
Why contextual drill-down and preserved evidence make interactive investigation more useful than static reporting.
Domain Knowledge Is Part of the Software
Manufacturing applications become more useful when production codes are translated into explicit, conservative, and auditable domain models.
Building a Digital Twin Before Building AI
Why simulation fidelity depends on manufacturing knowledge, classification, and production context before machine learning is introduced.
Dashboards Should Explain, Not Just Display
Using the Evidence Workspace as an example of designing dashboards around investigation, context, and traceable production evidence.
What Real Manufacturing Data Teaches About AI
Lessons from production data that rarely behaves like clean benchmark datasets and often requires process understanding before machine learning.
Why We Stopped Building Reinforcement Learning First
A reflection on postponing reinforcement learning until the manufacturing environment became credible enough to support useful agent behavior.
Realism Before Optimization
Why Cutting Table RL is delaying timing models and reinforcement learning until real manufacturing data is available.
Simulation Before Reinforcement Learning
Why Cutting Table RL starts by modeling the manufacturing world before introducing agents or reinforcement learning.
OCR-First Rack Scanning for Bolt Finder
How Bolt Finder moved from barcode-first assumptions to a working OCR-first rack scanning pipeline.
OCR vs Barcode Detection in Real Factory Photos
A short note on why label readability and image detail matter before choosing the recognition method.