Eletivo

A campaign operating system for Brazilian elections

Company: Eletivo
Role: Co-founder & CTO
Timeline: Jun – Oct 2026
Website: eletivo.app

Product design
Brand
Front end
Back end
AI

Results

products on one data layer, designed and built by one person: 4; products on one data layer, designed and built by one person
votes reconciled between the raw files and the map: 1B+; votes reconciled between the raw files and the map
of 2022’s proportional seats reproduced by the seat estimate: 95.4%; of 2022’s proportional seats reproduced by the seat estimate
tools behind the AI analyst, from vote maps to tasks and press: 45; tools behind the AI analyst, from vote maps to tasks and press

The project

Eletivo brings electoral data, street operations, social media and press into one place, for everyone from a campaign’s command to its agents on the street. As co-founder and CTO, I designed and built it end to end, essentially on my own with AI agents: the brand, the data pipeline and four products on one data layer.

01TSE open dataVotes by polling section, polling places and the electorate, published as raw CSV files.

02Validation gateA state goes live only when the votes in the raw files equal the votes served.

03One data layerA read layer shaped for the map, split by state, beside the campaign’s own teams, tasks and evidence.

Siteeletivo.appWhere a campaign finds Eletivo and picks its license.Chapter 5
PlataformaThe command centerThe votes map, tasks and the AI analyst.Chapters 1, 2 and 4
CampoThe street team’s phoneTasks, check-ins and interviews, offline first.Chapters 2 and 3
AdminThe back officeClients, licenses and leads, every change audited.Chapter 5

The problem

Brazilian campaigns still run on spreadsheets and WhatsApp groups. The electoral court publishes every vote down to the polling section, but as raw files. Coordinators plan in one place, street teams report in another, and nobody can check what actually happened on the street.

01Coordinators can’t see where their votes are: the electoral court publishes them as millions of raw rows, not a map.See how it was solved
02Work is handed out in group chats, and nobody can tell what was really done on the street, or prove it.See how it was solved
03Field agents work on foot, in the sun and often with no signal, on tools made for a desk.See how it was solved
04Every question about the data waits for an analyst, and general chatbots make numbers up.See how it was solved
05And for the product itself: four apps for very different people, built by one person with AI agents, had to work as one system.See how it was solved

How I solved it

One answer per pain, with the evidence behind each.

Every vote on one map, down to the polling section

Coordinators need to know where their votes are, and the electoral court’s raw files can’t tell them. I built the pipeline that turns those files into a read layer made for the map: a candidate’s whole state in one view, then city, polling place and a single section. The map has three readings, absolute, nominal and proportional, and caps its color scale so a few outliers don’t wash out the rest.

From the platform demo on eletivo.app: São Paulo’s 2022 race for governor, from the state to a polling place, then the proportional reading.

Votes map: Tocantins, 2022 governor race

Votes map zoomed into Palmas with a polling place open

A whole state at once

Tocantins’ 2022 race for governor: 481,496 votes across 139 cities, with the ranking of cities beside the map.

01 / 05

A state goes live only when every vote matches

Four elections in 10 states, every office and round, streamed from the raw files and checked on the way in. Querying only a state’s own slice of the data took the city list from about 10 seconds to tens of milliseconds.

Raw rows in, rows served, votes matched

Raw rows Rows served

Rows and votes per state, from the import validation reports
State	Raw rows and rows served	Votes matched
SP	51.6M raw, 14.1M served, 3.66× smaller	447,007,384 votes, all matched
MG	19.2M raw, 6.2M served, 3.11× smaller	199,189,102 votes, all matched
RS	10.8M raw, 4.0M served, 2.69× smaller	119,074,574 votes, all matched
PR	10.2M raw, 3.1M served, 3.33× smaller	103,234,999 votes, all matched
SC	6.1M raw, 1.7M served, 3.50× smaller	74,848,835 votes, all matched
MT	2.7M raw, 0.7M served, 3.61× smaller	30,130,286 votes, all matched
MS	2.6M raw, 0.7M served, 3.97× smaller	26,649,800 votes, all matched

Rows: 103.3M to 30.5M; 3.38× fewer for the map to read
Votes: 1,000,134,980; matched between raw and served
State-years: 28 of 28; passed the gate: 7 states × 2018, 2020, 2022, 2024

From the validation reports of two import runs on GitHub Actions (30138595939 and 30180179627). ES, GO and TO were imported too, but their logs weren’t kept. Votes add up four elections, every office and both rounds: they are votes, not voters.

A seat estimate that knows when to stay quiet

On top of the data, a seat estimate replays 2022’s proportional races. Where its replay disagrees with the official result, it refuses to say how many votes it takes to win.

681 of 71495.4% of 2022’s seats reproduced

Reproduced Missed

20 races, state and federal deputies in 10 states

Choose one to see its seats.

From the log of the estimate’s replay run on GitHub Actions (31631875004).

Every task has an owner, a deadline and proof

Group chats lose track of who was asked to do what. In Eletivo, every task has an owner, a deadline and the proof it needs: coordination assigns it on the platform, the agent does it in Campo, and the evidence comes back to the same task.

eletivo.app on an iPad: Decida na Plataforma. Execute no Campo. — How eletivo.app puts it: decide in Plataforma, execute in Campo, and the task comes back with its evidence.

Proof by kind of action, not a rule builder

Each action on Campo’s home carries its own evidence. Coordinators can override a check, and every override is audited.

Check-in
At an event
Only counts inside the event’s radius, 200 m unless the coordinator sets another.
Visita
A door-to-door visit
How the door went, talked, no answer or refused, and where.
Eleitor
Registering a voter
Name and phone with the GPS of the spot; a phone already on file is flagged.
Entrevista
A survey interview
The answers, with the audio recorded unless the agent turns it off.

Every task by status

Planned, in progress, blocked and done, each with its kind of action, time window and progress.

01 / 03

Each person sees their own part of the campaign, and the screen says which: the whole campaign, their coordinations, or their direct team.
A completion whose photo was lost can still be accepted: a retry will never produce the photo.

The campaign as it is organized

Command, coordinations, field leaders and agents, each with a role that decides what they can see and do.

Coordinations, with the headquarters being placed on the map

The cockpit

Coverage by area, pledged votes, field productivity and risk, with the areas ranked by priority.

01 / 06

A field app for the street, not the desk

Agents work on foot, in the sun, often with no signal. Campo is an offline-first app: every action is saved on the phone first and sent when the network returns. A native version was built and cut; Campo installs from the browser, with no app store in between.

9:41

Three quick actions an agent repeats all day, the open tasks, and anything still waiting to be sent.

The leader’s area

Field leaders run a crew from the same app: their coordination’s agents, tasks, what was recorded today and what needs attention. On a tablet it opens up to the width of the screen.

Gestão on a phone: the leader’s overview

Agents, tasks, today’s records and what’s late, for one coordination.

Campo running locally against the project’s showcase seed (fictional agents in Tocantins), signed in as a field agent and a field leader.

What the street sends back

The same work, as the platform shows it to coordination.

The operational map

Areas, visits, agents and registered voters in one city, block by block.

01 / 04

Field data from the campaign’s showcase seed.

Light only
So it stays readable in direct sun.
Big titles to browse, small ones to act
Large titles on lists, compact ones on forms, checked side by side with Apple Health.
Honest failures
Anything that can be resent waits in the outbox; lost evidence is marked instead of retried forever.
A position, not a trail
The leader’s map shows each agent’s last known position, with an accessible list beside it. Agents are never ranked.

An AI analyst whose numbers come from the data

A question about the data usually waits for an analyst, and a general chatbot will make up a vote count. In the Oráculo, the model plans and narrates; every number and every card comes from 45 tools on the server.

Who does what in an answer

The model
The server
01The modelPlansReads the question and picks which tools to call.
02The serverRuns the tools45 of them, deterministic, over the electoral data and the campaign’s own.
03The modelPicks a cardChooses a source and a layout from fixed lists. It can’t type a number.
04The serverFills the cardEvery row comes from what the tools returned.
05The modelNarratesWrites the answer from the tools’ output, with no tools left to call.
06The serverAdds the noticeThe label the electoral court requires on AI content.

What the tools cover

Election results
2026 candidates
Electorate by area
Tasks
Agenda
Social
Press

The AI proposes, people decide
It can draft a task, but nothing is created until someone confirms it.
The notice isn’t optional
The legal label on AI content is added by the system, so the model can’t leave it out.

Four products that work as one

Each product has its own user and setting: a desk, a phone in the sun, a back office, a sales page. Built by one person with AI agents, four apps could easily drift apart. I wrote the design system down first, as rules the agents had to follow, and let each product speak its own dialect on top of it.

One brand

One accent

#001EFF

With the dither atmosphere every product’s sign-in shares.

Space Grotesk, for headings
Decida na Plataforma. Execute no Campo.
JetBrains Mono, for figures and labels
R$ 9.900,00

Candidate
Command
Strategist
Coordinator
Field agent
Interviewer
Logistics
Territory
Legal
Communication
Social
Press

The campaign dashboard’s icons, engraved in wireframe and checked on the three surfaces they live on: light, brand blue and dark.

Rules the agents can’t skip

Written as tests
At least 4.5:1 contrast in dark mode, and surfaces that get lighter as they rise, in the token tests.
Enforced by lint
The surface hierarchy is a lint rule, not a review comment.
Deliberate dialects
The map is always dark, Campo is always light, and the back office speaks the platform’s language.
One set of tokens
Shared tokens and components across the four products, with the dark theme derived from the map.

The site: one price per office

An audit of the pricing page found no h1, three cards that looked alike and buttons hidden from screen readers. Campaigns don’t choose between plans; the office they run for sets the price. So the page became one license per office, picked by tab.

eletivo.app pricing page (Before the audit)

The back office: friction follows consequence

Anything reversible asks for a reason, and the reason becomes the audit record. Deleting a client is for owners only and means typing its exact name.

Outcomes & learnings

Paused until Brazil’s 2028 municipal elections, with the site still live at eletivo.app.
Electoral data a campaign can act on, where nothing reaches the map until it matches the raw files.
An AI analyst that answers with the campaign’s own data and can’t put a made-up number in a card.
The AI phone-survey channel was built and tested, and stays pre-pilot.

What I learned

With AI agents, consistency comes from written rules, not from review. I wrote the design system down first, and turned each repeated mistake into a rule or a lint check.

Cutting is design work too. Each of these pulled away from what Eletivo is for: official data and proof of field work.

Built, then cut

Campaign finance and contracts−7,272 lines
WhatsApp messaging−15,646 lines
WhatsApp surveys−9,689 lines
A native app, three timesabout −32,000 lines
Political news−2,086 lines
Parliamentary amendments7 days in the product

Book a call Back to all work

Next case

Onzi

From App Review rejection to the App Store