Journal

Field notes

Thesis, playbooks and engineering notes on making AI work for the companies that build and run the physical world.

Featured · Thesis

The physical world runs on drawings

Every building, plant, pipeline and grid is built and changed through documents no machine can read. Why that stayed true, why chat AI won't change it, and what we are building instead.

Operon Team · Oct 14, 2026 · 9 min read

Read the post →

All posts

19 posts

Private EquityOct 14, 20268 min read

The missing layer in the industrial roll-up

Buy-and-build in industrial services tends to stall where the work lives: in each add-on's estimates, drawings and tribal knowledge. A shared intelligence layer is how a portfolio starts working as one.

PlaybookOct 14, 20268 min read

Bid on counts, not allowances

Under a bid deadline, the parts of a mechanical estimate nobody had time to count turn into allowances. A verified AI take-off puts counts back in their place, each one traceable to the sheet it came from.

ThesisOct 15, 20268 min read

One layer, four industries

A mechanical estimator, an EPC document controller, an LDAR coordinator and a process safety engineer do different jobs from the same kind of document. Why the reading problem is shared, and why one verified layer can serve all four.

Private EquityOct 15, 20267 min read

Finding add-ons before the banker calls

Many of the industrial services companies a platform could buy will never run a formal sale process. Sourcing them means reading the market directly: licenses, permits, project records and the companies' own work. A playbook for an add-on pipeline built from data.

EngineeringOct 14, 20269 min read

Reading a P&ID like an engineer

Symbol sets, tag grammar, line tracing across sheets, and why OCR plus object detection falls short. A technical walk through how a P&ID becomes a typed graph, and how revision deltas fall out of it.

Private EquityOct 14, 20268 min read

Diligence on the drawings

A contractor's bid archive, drawing sets, change orders and close-out files are the most honest record of how it works. What a sponsor can learn from them before signing, how to get access, and how the findings shape the first 100 days.

PlaybookOct 14, 20267 min read

Submittal review without the stack

Mechanical submittals are checked by reading product data against Division 23 and the equipment schedule, page by page. An AI-assisted check makes that comparison line by line and drafts the comments. The project engineer still decides what goes back.

EngineeringOct 14, 20267 min read

Why we verify every tag

Detection accuracy above 97 percent still leaves errors scattered through a drawing set. Here is how we route uncertainty to engineers, record who signed off, and keep raw model output out of your systems.

PlaybookOct 15, 20267 min read

The inbox is a workflow

In an industrial services company, much of the real work arrives by email: bid invitations, addenda, RFIs, submittal returns, change requests. Treating the inbox as a set of workflows, each ending in a draft a person approves, is where back-office automation starts.

Private EquityOct 14, 20267 min read

The first 100 days: bringing an add-on onto a shared AI layer

A phase-by-phase playbook for bringing a newly acquired contractor onto a shared AI layer: who owns what, what to watch, and the failure modes that stall it.

EngineeringOct 15, 20268 min read

Reading a mechanical set like an estimator

Legend sheets, schedules, duct plans, enlarged plans and the notes that change everything: how an HVAC estimator reads a mechanical set, why it is hard for a machine, and how a verified take-off is built from it.

Field NotesOct 14, 20267 min read

Turnover packages your owner can query

At close-out, everything a project learned about the asset is handed over as thousands of PDFs. Built as a verified register indexed by tag, the turnover package can be checked by query instead of by reading, and it keeps working after the project team leaves.