Reid's first call on the new site. Compute: rent cloud machines first and buy once real numbers are in. Reid goes ahead with the Ector County OCR run, and one app will hold the meeting notes, the intake and the inventory. The walk-through found several page problems to fix.
What this call produced 19
To-do T58: Bring everything into one app: Reid's Cletus rig and ruling tools, the inventory's explore view, and this meeting app
To-do T57: Fix what the walk-through found: a typed question vanished after send; the ask box looks like a name box; the page did not show Reid as signed in; the read-back post was hard to find; any talk during a yes-or-no question counts as the answer; there is no way to share a link
Decision D19: One app does everything: Reid's intake app joins the meeting notes and the inventory view in one app, with each part in its own code repository behind one entry point
To-do T54: Put the records into the sandbox database and define the actions from the transfer automation plan, using plain code first and an AI setup only where code falls short
Not filed yet
To-do T53: Run the OCR on the Ector County records
Decision D17: The intake in the Command Center is the old first version and is retired; the one we build on is in the deal data project
Not filed yet
To-do T52: Meet again tomorrow: wearables, and getting the environment set up so we can start processing records (tidy-up a week later)
Not filed yet
9 older
To-do T51: Get Joe's answers to Reid's questions to Reid; they were given to Claude but never reached him
Not filed yet
To-do T50: Finish the answer sheet for the intake app tests, so the setups on the Cletus rig are scored for accuracy, not only for agreeing with each other
Not filed yet
To-do T49: Move the meeting notes app into its own code repository
To-do T48: Pick two hardware options: one for day-to-day work, one for peak loads like the big county data sets. Include the point where renting costs as much as owning, and that renting means no upkeep
Rent cloud machines first and buy once the numbers are in; Reid sets up the test
Done when: We agree the next step on compute and who takes it · Housekeeping
What do we do with the Ector and Martin County records?
Partly met
Reid runs OCR on the Ector County records now; how the records get used after that is not planned yet
Done when: We agree how the two counties' records get used and who does it · Housekeeping
How do we proceed this week on building things out?
Partly met
Each of us has tasks with owners, but most have no day yet
Done when: Each of us has this week's tasks with an owner and a day · Housekeeping
Full close-out
Not decided yet
How the Ector and Martin County records get used after the OCR run, because Reid wants to see the OCR results and the Title Lab comparison first. Waiting on: Reid
Which hardware to buy, because the rented cloud test has to give real numbers first. Waiting on: Reid's cloud test, then Joe
To-dos by person
No owner yet: Get Joe's answers to Reid's questions to Reid; they were given to Claude but never reached him
No owner yet: Meet again tomorrow: wearables, and getting the environment set up so we can start processing records (tidy-up a week later)
Claude (owner unconfirmed): Fix what the walk-through found: a typed question vanished after send; the ask box looks like a name box; the page did not show Reid as signed in; the read-back post was hard to find; any talk during a yes-or-no question counts as the answer; there is no way to share a link
Claude (owner unconfirmed): Bring everything into one app: Reid's Cletus rig and ruling tools, the inventory's explore view, and this meeting app
Claude: After this call, work with Reid to run the meeting app from the cloud server Reid set up last week, so anyone can use it. Next: Get the server's address and access from Reid, by after today's call
Claude: Move the meeting notes app into its own code repository
Joe: Make the hardware decision
Joe: Pick two hardware options: one for day-to-day work, one for peak loads like the big county data sets. Include the point where renting costs as much as owning, and that renting means no upkeep
Reid: Set up the rented cloud test and bring back the numbers
Reid: Finish the answer sheet for the intake app tests, so the setups on the Cletus rig are scored for accuracy, not only for agreeing with each other
Reid: Run the OCR on the Ector County records
Reid: Put the records into the sandbox database and define the actions from the transfer automation plan, using plain code first and an AI setup only where code falls short
Reid: Once Title Lab access comes through, compare how it handles the data with ours
Reid: Send Joe a write-up after the call: what we discussed, what is done, what is left, and the plan for the repositories, by today
Topics carried forward
AI server hardware
Reading old county records: OCR plus a vision model
Nothing in: Which part of the vision moved, Project stage changes, Drafts waiting for review.
Claude's replies · 12
Read-back
Read-back. Say "all right" or press it if both lists are right; otherwise say which item is wrong.
Decisions
The Command Center intake is the old first version and is retired; we build on the one in the deal data project.
Reid goes ahead with the OCR run on the Ector County records now.
One app does everything: Reid's intake app joins the meeting notes and the inventory view, each part in its own code repository behind one entry point.
To-dos
Joe: make the hardware decision; pick two hardware setups (day-to-day and peak), with the point where renting costs as much as owning.
Reid: set up the rented cloud test; run the OCR on Ector County; finish the intake test answer sheet; load the records into the test database and set up the transfer-plan steps; compare with Title Lab once access comes; send Joe today's write-up, today.
Claude: after this call, work with Reid to run the app from his cloud server; move the app into its own code repository.
No owner yet: get Joe's answers to Reid's questions over to him; set tomorrow's call (wearables and setting up the environment).
Captured, and what's next
Captured: Joe picks two hardware setups, one for day-to-day work and one for peak loads like the big county data sets, and finds the point where renting costs as much as owning. Rented machines also need no upkeep.
Next on the agenda: how we proceed this week. For each task this week, say who does it and by what day.
Glasses, drones or a service for scanning mail
Short answer: for mail, a mail-scanning service or a sheet-fed scanner beats glasses and drones today. The glasses can work as a hands-free camera, but only through an app we would build ourselves. Drones don't fit indoor paperwork.
The options
Option
How it works
Cost
Fit for us
Virtual mailroom service
Mail goes to their address. They open it, scan every page, and send the files the same day.
about $10 to $50 a month for small offices, more for many recipients
Best first step. No hardware, no staff time. The files then go into our usual reading process.
Sheet-fed office scanner
Someone opens the mail and drops the stack in. It scans both sides.
a few hundred dollars
Good if mail has to stay in the office. One person, a few minutes a day.
Meta glasses
Our own phone app takes the glasses' camera pictures as someone looks at each page.
the glasses plus a small custom app
Possible, not ready. Meta's camera access is still a preview for developers, apps can't be published to stores yet, and there's no built-in document scanning.
Envelope-opening machines (OPEX)
Open, empty and scan thousands of envelopes an hour.
quote only; made for large mailrooms
Too big for us. The mail services above run these machines for us.
Drones
No real product for paper mail.
-
Not a fit.
Where the glasses could help: in the field, like a courthouse visit where someone reads old books and can't use a scanner. That is a separate test from mail.
Sources: Meta's developer toolkit coverage (Road to VR, UploadVR, May 2026 display update), PostScan Mail and Alliance Virtual Offices price pages, OPEX product pages.
Testing a 4-card Dell in the cloud
Short answer: rent Amazon's 4-card machine (g7e.24xlarge) for a day. It is the closest match to a 4-card Dell Precision 7960, at about $17 an hour, or about $6 an hour if you accept that Amazon can take it back mid-run.
Ways to test it
Option
What you get
Price per hour
Good for
Amazon g7e.24xlarge
4 RTX PRO 6000 Blackwell cards (96 GB each), 96 CPU cores, 1 TB memory
about $16.57, or about $5.55 interruptible
The real test. It stays inside Amazon, next to Bedrock and your records.
Amazon g7e.12xlarge
2 of the same cards
about $8.29
A cheaper first run, or the 2-card Dell question.
RunPod, Vast.ai and similar
The same card, rented one at a time
about $1.64 to $2.49 per card
A quick, cheap check that a model fits and runs. Moving data there is extra work.
Three things to know
Close, not identical. The cloud cards are the server version. The Dell takes the workstation version, which runs on less power, so the cloud will read a little faster than the Dell.
Ask Amazon first. New accounts can't start a 4-card machine until Amazon raises the limit for GPU machines. That request can take a day or more, so file it today.
Don't resize the Cletus box. If "Cletus Rig" is the always-on meeting box, changing its machine type replaces it and wipes its signed-in setup. Start the GPU machine as a separate one, and shut it down when the test ends.
A day of testing on the 4-card machine is about $130 to $200, against a card price of about $16,000 each.
Prices come from public price trackers this month (Thunder Compute, runs-on, ComputePrices). Check Amazon's own price page for your region before you start.
Before we move on
Before we move on, three quick things:
Say yes to decision 1 now: the Martin and Ector County records are a test of how far we can get, and they do not change the hardware choice for now. Yes or no?
Then decision 2: test on rented cloud machines first, and use those numbers to decide what to buy. I'll ask right after.
Who sets up the rented cloud test? No owner yet.
Then we move to: how we proceed this week.
How our name-recognition process works (corrected)
Short answer: a deal's documents go in, a filled-in intake form comes out, with no paid AI along the way. It replaced the old setup that paid Gemini to find values and Claude to check them.
How it runs, step by step
Read the documents. Every file in the deal folder is turned into plain text.
Tag the values. A free name-recognition model running on our own computer marks the names, tracts, leases, wells, decimals and dates in that text. A set of fixed rules catches what it misses.
Read the allocation sheet. The seller's Excel table gives interests, tracts, decimals and leases.
Connect the pieces. Tracts are matched by county, survey, section, block and abstract. Wells are tied to their leases and tracts. Pooled units are split back to tracts.
Fill the form. Fixed rules work out the totals, purchase price, commission and effective date, then write the intake form.
Grade it. Each form is checked against an answer sheet a person made for that deal.
How well the tagging step did (tested on three deals)
Deal
Values found
Of what could be found
Sandra Lazarone
73%
96%
Leo H. Thomas
89%
98%
Saxet Energy
75%
84%
Where it stands. The tagging step is settled. Turning tags into a finished form is still being worked on: the bar is 50 of 60 fields right on the Lazarone deal. There is also a newer, separate test lane where an AI agent fills the form instead, graded against the same answer sheets.
Reid: this sits beside your OCR and vision tests. Step 1 already reads scanned pages with a free OCR tool when a page has no text, so your tests would compare against that.
How our name-recognition process works
Short answer: a deal's documents go in, a filled-in intake form comes out, with no paid AI along the way. It replaced the old setup that paid Gemini to find values and Claude to check them.
How it runs, step by step
Read the documents. Every file in the deal folder is turned into plain text.
Tag the values. A free name-recognition model running on our own computer marks the names, tracts, leases, wells, decimals and dates in that text. A set of fixed rules catches what it misses.
Read the allocation sheet. The seller's Excel table gives interests, tracts, decimals and leases.
Connect the pieces. Tracts are matched by county, survey, section, block and abstract. Wells are tied to their leases and tracts. Pooled units are split back to tracts.
Fill the form. Fixed rules work out the totals, purchase price, commission and effective date, then write the intake form.
Grade it. Each form is checked against an answer sheet a person made for that deal.
How well the tagging step did (tested on three deals)
Deal
Values found
Of what could be found
Sandra Lazarone
73%
96%
Leo H. Thomas
89%
98%
Saxet Energy
75%
84%
Where it stands. The tagging step is settled. Turning tags into a finished form is still being worked on: the bar is 50 of 60 fields right on the Lazarone deal. There is also a newer, separate test lane where an AI agent fills the form instead, graded against the same answer sheets.
Reid: this sits beside your OCR and vision tests. It starts from text the documents already carry, so a scanned page with no text needs your OCR step first.
Ready to start
Nothing else needed. Moving to item 2: compute, 20 minutes.
Noted from the walk-through, to fix later:
Reid's typed question vanished after he pressed send.
Answers to typed questions land in a place that is hard to find.
The ask box looked like it wanted a name.
The latest chat
I don't see a new chat message yet. The last thing typed on the page was Reid answering "who is here".
Reid: type your question in the Chat box and press send. It should show up within a few seconds, with an answer under it.
Proposed agenda
Next: agree the agenda. Here is a proposal for the 90 minutes:
App walk-through for Reid, 10 minutes
Compute: what we buy or rent next (G1), 20 minutes
Ector and Martin County records (G2), 25 minutes
How we proceed this week (G3), 20 minutes
Wrap-up: read back decisions and to-dos, 5 minutes
Say yes, or change the order or the minutes.
Getting started
Who is here?
How long do we have?
What do we want to walk out with?
First 10 minutes: Joe walks Reid through the app. Answer these out loud after the hellos.
Listening
Claude is listening. Say "Hey Claude, ..." in the call and the reply appears here.