Jefferson Alves
In production since May 2026 · Lisbon

I build AI operating systems for businesses: one system for an operation that used to live across tools that never talked to each other.

Orkestra is the system. It has run a Lisbon tour operator’s day, every day, since May 2026: bookings, schedule, dispatch, messages, money. I built it, I work there, and I measure what it gives back.

See the system running →
Hire me on Upwork
20 minutes · Google Meet · bring the jump that costs the most
Jefferson Alves
Jefferson AlvesAI Systems Architect · Founder, Orkestra · Lisbon
The dispatch flow, as it runs · demo tenant · 12 s loop, no soundIn production since May 2026 · Lisbon
≈ 20 h / month
of manual work removed, across messaging, dispatch, the weekly schedule and the books
August 2026 · this operator only
414 messages
sent to customers in six languages, 404 delivered
Message log · May to Sep 2026
6,499 bookings
read from the calendar once, never retyped
Calendar sync · history since 2022
Measured on production data, one tenant, no demo data. Not a projection.How each number was measured ↓
The case · Live Portugal, Lisbon

Five times a day, a booking was retyped into ChatGPT.

A tour company in Lisbon used to build its week by hand in a spreadsheet, export it as a picture, and post it in the drivers’ WhatsApp group. Bookings lived in a calendar. The accounts lived in other spreadsheets. And customer reminders were typed one at a time, in six languages.

Today one system runs that day. The managers still decide. The machine prepares. Nothing reaches a driver before the manager dispatches, and no message leaves without a person tapping send.

Full disclosure: I work at that company. I built the system that runs the operation where I work, and I operate it every day, so I live with my own decisions.

Walkthrough 1 of 4 · the messages · demo tenantRecorded Sep 2026
Measured task by task

Four tasks, four before-and-afters.

I measured the operation task by task. Each one has its own number, and the four add up to the twenty hours.

01Customer messages

The messages write themselves. Sending stays one tap.

≈ 6 h
a month, back
111 messages a month
Before3.5 min per message, by hand

Copy the booking out of the calendar into ChatGPT: name, pax, language, pick-up, time. Pick the right template out of a dozen. Open WhatsApp, find the number, paste, send. Then again for yesterday’s customers, to ask for a review.

Now0.2 min to review and send

The system reads the calendar and writes each message in the client’s language: the reminder the day before, the review request the day after. They wait in a queue. A person releases each one, free to change it or not send it at all.

How it was measured: 111 reminders and review requests prepared in August 2026. 3.5 min by hand, 0.2 min to review and send. Source: the message log of this operator, one tenant, no demo data.

The send queue: reminders, review requests and driver briefings, each in the client’s language, each sent with one tap.
The send queue: reminders, review requests and driver briefings, each in the client’s language, each sent with one tap.
02The weekly schedule

The schedule stopped being a picture and became data.

≈ 3 h
a month, back
built by the managers, published in one click
Beforea spreadsheet, exported as a picture

Built in a spreadsheet, exported as an image, posted in the WhatsApp group. And there it stopped: a picture is a dead end, nothing downstream can read it. Every assignment meant going back to check who was working.

Nowone click, to the group and to each driver’s app

The managers still build the week; they know who works well with whom and who takes the German tour. Now they build it on a screen made for the job, and one click publishes it to the group and to each driver’s app.

How it was measured: The hours come from no longer exporting and sharing by hand, and above all from no longer re-checking the schedule at each of the 128 assignments a month. Source: assignment log, August 2026.

Walkthrough 2 of 4 · the scheduleRecorded Sep 2026
Drivers by row, days by column, tuks dragged into cells.
Drivers by row, days by column, tuks dragged into cells.
One click: the week goes out as an image to the group and into each driver’s app.
One click: the week goes out as an image to the group and into each driver’s app.
03Dispatch and the driver’s app

A last-minute tour used to mean five phone calls.

≈ 4 h
a month, back
128 briefings a month
Beforecall the drivers one by one, mid-tour

Open the schedule to see who was working. Work out who already had a tour at that hour. Call drivers one by one, out in the street, to find out who was actually free. Then type the briefing.

Nowone screen, one click, the driver’s phone buzzes

The system has read the week the managers built, who is out on the street, who is on holiday, which languages each driver speaks. It shows who can take this one. One click assigns, and the tour lands in the driver’s own app.

How it was measured: 128 driver briefings prepared in August 2026. 2.0 min by hand, 0.2 min now. Source: dispatch log, this operator only.

Walkthrough 3 of 4 · dispatchRecorded Sep 2026
Only the drivers rostered that day, ranked by language and availability.
Only the drivers rostered that day, ranked by language and availability.
The driver’s app: today, tomorrow, the week, and the balance. Each driver sees only their own work.
04The money

The balance settles as each tour closes.

≈ 6 h
a month, back
month-end close: 3 h 30 to 30 min
Beforespreadsheets, updated by hand

Each driver’s balance lived in a spreadsheet, updated by hand. Each partner hotel’s commission was worked out at the end of the month, from the calendar and another spreadsheet. The close took an afternoon.

Nowsettled at the moment the tour ends

What the driver received and what he is owed, settled the moment the tour ends. Each partner’s commission ready when the month ends, from the same data that ran the day. This layer is plain deterministic code, because it has to be right.

How it was measured: 99 balance movements in 31 days, and one month-end close, 3 h 30 by hand to 30 min. Source: the accounts ledger, August 2026.

Walkthrough 4 of 4 · the booksRecorded Sep 2026
A driver’s month: bookings, expenses reimbursed, balance, and the movements behind it.
A driver’s month: bookings, expenses reimbursed, balance, and the movements behind it.
A partner hotel’s month: tours, total value, commission, net. Ready to download, nothing re-entered.
A partner hotel’s month: tours, total value, commission, net. Ready to download, nothing re-entered.

Nothing is typed twice.

The biggest cost was never one task. It was the operation living in tools that did not talk to each other. Every booking is read once. The same data that writes the customer’s message later tells the system which driver can take the tour and in which language. The schedule the managers build is what the assignment picker reads. And when the driver closes the tour in his app, his balance closes with it.

Every booking in one list, synced from the calendar: time, tour, pax, customer and channel, driver and vehicle, status.
Every booking in one list, synced from the calendar: time, tour, pax, customer and channel, driver and vehicle, status.
Ana Paula
Ana Paula
Founder and manager, Live Portugal
In her words

Jeff’s work had a very significant impact on our operation. The tools he developed let us raise productivity, cut the time spent on operational tasks considerably, and have far more control over the company’s different processes.

Beyond the technical quality of the work, I would also highlight his professionalism, his ability to understand the needs of the business and, above all, his readiness and availability to find solutions and answer the challenges that came up along the way.

I recommend Jeff without hesitation to any company looking for someone competent, responsible and focused on creating solutions that truly generate value for the business.

Written in Portuguese · translated
The other half of the job

Where AI does not touch.

Writing this list with the client is part of the delivery, not an appendix. A team that does not know where the machine stops distrusts it entirely, and the system dies of disuse in three months. The default:

  1. 01
    Decisions about peopleHiring, letting go, evaluating, changing someone’s shift.
  2. 02
    Money going outPayments, the final price, discounts, credit.
  3. 03
    Messages that leave with the company’s name, unreviewedWhen the subject is a complaint, a cancellation, money, or something that already went wrong.
  4. 04
    Clinical, legal or tax judgmentA model can draft. It does not sign.
  5. 05
    Any output whose data source cannot be tracedIf a number has no source, it does not go out. Missing data pauses the flow and tells a person.
  6. 06
    What the team already does well and fastAutomating that does not pay for its own upkeep.

In Live Portugal that line is the whole design. The schedule is built by the managers because it is a judgment about people. The balances are plain code because they must be right. The customer send is a human decision because the cost of being wrong is the channel the business runs on.

How it starts

The map and the first module.

The same way every time, in this order. Skipping a step is the most expensive mistake, and it is almost always the second one.

Week 1

Map the operation. Time the costliest jump.

Every tool the operation lives in, and every copy between them. Minutes per task, ten real runs, two different people, before anything changes. Without a baseline there is no project.

Weeks 1 to 2

Draw the line. Build the first module.

What the AI will do and what it will never do, written and approved before the build. The first module is built so the second one fits. The exception goes to a person, visibly.

Day 30

Measure again, same unit.

Same task, same method. You keep the map of your operation and one module running, whether or not you continue with me.

Fixed price, one to two weeks, quoted after the call. Best fit: you run bookings or jobs, a team in the field, and money that has to reconcile, and the operation lives in three or more tools that never talk to each other.

What you get, in full →

Who is behind this

Jefferson Alves.

AI systems architect in Lisbon, founder of Orkestra. Twelve years in music first. Then Managing Director of an eighty-person marketing agency across two offices in Brazil, which is where I learned the thing that decides these projects: a technically correct tool the team refuses is worth nothing. Then three years guiding tuk-tuk tours in Lisbon, for the company whose system I later built.

I am not an AI enthusiast. Half the job is saying where AI does not belong. I build with n8n, the Claude API, Supabase, Next.js and TypeScript, and I stay after launch: train the team, own the result.

I won’t call it done until it runs like music.

Music · twelve years
Music · twelve years
Brazil · eighty people, two offices
Brazil · eighty people, two offices
Lisbon · guide first, engineer second
Lisbon · guide first, engineer second
One conversation

Tell me which jump between your tools costs the most.

I reply with a scoped first module and a fixed price.

Hire me on Upwork
20 minutes · Google Meet · bring the jump that costs the most