Tomorrow's roster, ready by 8pm
Every request read and checked against your rules, the whole day planned together. Your supervisors look it over, change what they want, and send.
Built by hand, every evening. WhatsApps, calls, passes, leave, calibration dates: mostly from memory.
Waiting for you at 8pm. On a made-up day: 17 jobs ready to approve, 6 to look at, 1 flagged with the reason.
- What would your evenings look like with it?
- How would it work, and how is it built?
- What would we do together, week by week?
Every screen in this presentation uses made-up data: names, sites and requests are all invented.
Today, the whole day rests on one experienced supervisor
It works because he holds it all in his head. The aim is to write that down, so it is there on his days off too.
The system recommends. Your supervisors approve.
- Reading every request into a job card
- Matching on the rules, then ranking
- A kit list for each person, calibration checked
- The 8pm review: approve, or change with a reason
- A WhatsApp message to each technician, sent by the supervisor
- A phone app for technicians
- Van routes
- An owner's view of the whole team (appendix A1)
- How much more work the team can take
- Billing reconciliation
A WhatsApp at 3:42pm becomes a checked job by 8pm
It arrives the way it does today.
Shorthand is normal: "tmr", "Lvl 2", "same as last week". Your clients change nothing.
Each coloured phrase lands in a field. Anything the message doesn't say is marked for a check.
Each person gets their own kit list, and nothing out of calibration goes out.
"Calibrated until" is the date each unit's calibration certificate runs out. If it runs out before the job ends, the unit is swapped for one from the store.
Prepared by the system. Sent by the supervisor.
Site, gate, report time, partner, and the actual units with their dates. No new app.
At 8pm, one screen: 31 jobs, 57 people, planned together
| Job | Need | Recommended | Status |
|---|
Refinery A, Jurong Island: radiography, a crew of 2. Only one licensed radiographer is free. Nobody goes out alone on radiography, so it flags the job: call the client, move it to Friday?
This is real custom software, and it fits around the tools you already use
Four ways a request comes in
A supervisor forwards the client's message to one dispatch number. One tap.
Sent or forwarded to a dispatch address, PDF work orders included.
After a call, the supervisor types or says a line.
A paste box for the odd one out.
Each piece, and what to check with you
Only the last column needs you: tell me where today is different.
| Piece | What it does | What it uses | Please check |
|---|---|---|---|
| WhatsApp in | Client requests reach the system | A dispatch number on WhatsApp's official business service | Do clients message a business number, or supervisors' personal phones and groups? Either works: personal chats are forwarded |
| Email in | Requests and PDF work orders | One dispatch inbox | Microsoft or Google email? Any client portals? |
| Calls in | A note after each call | A "new request" box; type or speak | Roughly how many requests come by phone? |
| Your records | Staff, sites, equipment | Your own Excel or Google Sheets, read each evening | Where do they live, and who keeps them up to date? |
| Reader | Message into a job card | Claude, Anthropic's AI model, an outside service reached over the internet | Any client whose contract limits outside services? |
| Rules | Who may go where | Written from your sheets and your senior supervisor | Which rules live only in people's heads? |
| Day planner | The whole day at once | Google's free planning software, OR-Tools | Nothing; this part is mine |
| Kit check | Every unit in calibration | Your equipment register | Is calibration tracked per unit and probe today? |
| The 8pm screen | Approve or change | A web page; phone or laptop | Do supervisors plan on a phone or a laptop? |
| Messages out | Each technician's job and kit | Prepared, then sent from the supervisor's WhatsApp | Do technicians all use WhatsApp? |
| Where it runs | The app and its records | A small cloud account in Master Scan's name, in Singapore | Microsoft or Google account preferred? |
| Evening summary | Did it run, what it planned, anything failing | A short report to you and me | Who else should get it? |
What goes into it
- The rules engine
- The whole-day planner
- Kit and calibration checks
- The replay test
Nobody goes out without a person's approval
A supervisor approves every assignment and sends every message.
A missing pass, an expired certificate or an out-of-date unit takes that person or item off the list.
Swap in anyone who passes the rules. Overriding a block is a separate step that records who and why.
Every plan, approval and change: who, what, when.
A short summary to you and me: it ran, what it planned, anything failing. Counts only.
Before anything goes live, it is tested on your own past weeks
Three numbers a day, a few minutes, on a sheet I set up.
Your last 2 to 4 weeks run through the system; differences go to your senior supervisor.
Weeks with it against weeks without.
The team sent T-021. The system picked T-033 (UT Level 2, in the site's usual team).
Would that have been right?
3 · How we'll know it worked (the KPIs)
| Measure | Counted by | Good looks like |
|---|---|---|
| Wrong allocations per week | Your tally | Down |
| Supervisor minutes a day on allocation | Your tally | Down |
| Missing-kit incidents | Your tally | Down |
| Plans approved without changes | The system's log | Up, as your reasons become rules |
You're my co-pilot: each week you get something to react to
- Your people by name on your screens; the AI step works with IDs.
- Client and site names are swapped for codes before the AI step wherever we know them, so it sees as little as possible.
- It stays in your account, used only for this, deleted whenever you ask.
- The AI service keeps nothing for long: deleted within 30 days2 and not used for training without permission3; we never give that permission.
- A one-page note on all of this before you share anything.
- Free. I'm doing this to take load off your supervisors, and I'll learn from it too.
- Runs on your tools. If it ever stops, your team carries on as today.
- We both watch it. An evening summary goes to you and me.
- You own it. Your setup and data are yours.
- Either of us can stop at any time.
Starting is small: counting can begin this week
- Start three simple counts. Wrong allocations, missing-kit incidents, minutes spent allocating. I'll set up the sheet.
- Share four things once you're happy with the data note: staff skills sheet, sites and what each requires, equipment register, two to four weeks of past requests with who was sent.
- Thirty minutes with your senior supervisor in week 1.
A few questions for you
Your answers decide how version 1 is wired. Notes stay on this laptop.
What you told me, and what happens next
What you told me
- 30 to 50 requests a day, 50 at the most, with no cut-off time. Most come to the WhatsApp Business number; email runs on Google Workspace; one client started sending through its own portal in the last 10 days. Calls go into a notebook, then into Excel.
- Three departments, and one supervisor handles 80 to 90% of the work. That is where the system helps most, and the risk when he is away.
- The rules live with that supervisor today. They are easy to write down, and your supervisors will help.
- Permits are set per project, not per client: two lines at the same site can need different permits. Expiries are tracked in Excel.
- Calibration is checked by hand against the equipment list, because the automatic reminders did not stick.
- Supervisors plan on a desktop, with the phone as a second option.
- The tools you have been shown needed constant data entry and never recommended who to send.
The first month, as we agreed
- Two weeks testing on your past requests, comparing the system's picks with who was actually sent.
- Then two weeks with your supervisors leading as today, and the system alongside offering alternatives. It moves to recommending first only when you are comfortable.
- Every week you get the three counts: wrong allocations, missing-kit incidents and minutes spent allocating.
- No fixed deadline. Your data stays in your own Google cloud account in Singapore, and client and site names are coded before the AI step.
Next steps
Ideas for later, for you and your supervisors
Illustrations on made-up data, to spark ideas. I'd build any of these only if you want them.
A1 · Your view as the owner
A2 · The supervisor's phone at 8pm
A3 · The plan as a board
Tap a person and only the jobs they're cleared for light up; tap a job to see who can take it and why the rest can't.
- Your figures Your own figures, from our conversation on 1 October (what you told me; not checked against your records)
- Confirmed Anthropic automatically deletes API inputs and outputs within 30 days of receipt or generation by default. (privacy.anthropic.com)
- Confirmed Data Anthropic retains from the API is not used for model training without the customer's express permission. (platform.claude.com)