Software & Tools / Technical index
Independent software projects for engineering analysis orchestration, local AI, verification, and bounded automation.
These projects support requirements, first-principles checks, repeatable execution, evidence capture, and honest failure states. Maturity and limitations are shown alongside capability.
Turn an engineering question into a reviewable analysis: requirements, first-principles checks, geometry, CFD or structural analysis where justified, and inspectable evidence.
Aero carries a declared engineering question through requirements, first-principles checks, geometry, appropriate numerical analysis, evidence evaluation, and reviewable disposition without treating an AI-generated answer as computational authority.
ForgeClaw and OCR97 remain useful companion experiments. They are deliberately secondary to Aero on this index.
ForgeClaw explores how an agent can act on a declared surface, preserve what happened, and require separate evidence before a result is accepted.
OCR97 treats extracted text as evidence with provenance and uncertainty rather than as a guaranteed transcription ready for silent downstream use.
Selected notes connect implementation decisions to engineering concerns: defined authority, observable execution, independent acceptance, and evidence that survives failure.
A reference roadmap for a teens-sized GX10 engineering pilot: Luna prepares traceable data and bounded implementations, a lead reasoning model designs and reviews, and deterministic gates retain authority.
Read ↗Find saved research projects, source snapshots, and implementation briefs without confusing planned work with released capabilities.
Read ↗A research starter for native Creo-to-NX feature reconstruction: pinned open-source repositories, known gaps, edit-behavior tests, and a bounded implementation brief.
Read ↗Public source downloads: Routing Lab, Engineering Validity Gate and Rev 2.3's controlled Physics-AI reference adapter—with measured results and explicit limitations.
Read ↗A real pretrained inference, a dimensional check, and a checkpoint-provenance audit. What executed, what disagreed, and why Aero keeps the result research-only.
Read ↗Five-repeat inference measurements, fresh-prefix controls, context scaling and honest limits. Faster responses are not yet faster engineering decisions.
Read ↗Why bounded agent workflows should define authority before acting, preserve execution evidence, and make acceptance a separate decision.
Read ↗A draft technical note on how missing rows, damaged labels, and uncertain text propagate into downstream tool decisions.
A draft measurement template for comparing local execution, cloud planning, latency, energy, and failure costs without cherry-picking.
ForgeClaw and OCR97 were originally published through Precision Arts Lab. That attribution is secondary here to their role in Alex's independent technical work.