Operação Serenata de Amor: open-source Rosie AI audits Brazil's congress expenses

Brazilian law makes congress members' reimbursement claims public, yet the Chamber receives more than 20,000 of them every month — too many for citizens, journalists or auditors to check by hand, which is exactly how abuse slipped through.

An open-source AI learns Brazil's reimbursement rules, scans every congress claim, and tweets suspect expenses to the public and the politicians themselves.

Data scientist Irio Musskopf crowdfunded roughly R$80,000 in September 2016, and a core team of about ten plus over 600 volunteers turned the written reimbursement rules into software. Rosie, an open-source machine-learning bot, flags anomalies such as overpriced meals or claims showing a member in two cities at once, has posted its suspicions on Twitter since 2017, and its companion platform Jarbas lets anyone browse and verify flagged receipts through a website and API.

By early 2018 more than 9,000 reimbursements had been flagged as suspicious; Rosie had tweeted 967 times by late 2019, when the group's own research found congress meal spending had fallen about 10 percent since her introduction. Coverage reached more than 90 media outlets, including Brazil's biggest — but enforcement stayed the weak link, with only dozens of politicians paying money back.

Brazilian law already gave citizens the raw material for accountability: every reimbursement claimed by a federal congress member is public. But the Chamber of Deputies processes more than 20,000 claims a month, and the data lived in dense official portals full of jargon — so in practice almost nobody audited it. Operação Serenata de Amor, begun in 2016, was an attempt to make that transparency readable and continuous rather than occasional.

The project was born from civic hackers who found the story irresistible: the name puns on the Swedish 'Toblerone affair' that ended a deputy prime minister's career over a chocolate bar on an official card, on a Brazilian candy called Serenata de Amor, and on the grand names of Brazilian federal police operations. Irio Musskopf shared the idea, a crowdfunding campaign raised roughly R$80,000 in September 2016, and a team of about ten people backed by more than 600 volunteers converted the reimbursement regulations into code.

The software became Rosie, an open-source machine-learning application that studies every claim and assigns it a probability of being irregular — flagging, for example, overpriced meals or receipts that place a congress member in two locations at the same time. Since 2017 Rosie has posted its suspicions on Twitter, addressing citizens and the accused politician alike, while Jarbas, a companion website and API, assembles the scattered official data into one searchable dashboard so people can verify or dismiss each alert themselves.

The numbers show a tool that worked as a deterrent rather than a prosecutor: by early 2018 more than 9,000 reimbursements had been flagged, Rosie had tweeted 967 times by late 2019, and the group's research found meal spending by congress members had dropped about 10 percent since her introduction. More than 90 media outlets covered the project, yet actual refunds stayed in the dozens — evidence that the harder problem was not detection but getting the institution to act on what a citizen-built robot had found.

It reversed the usual logic of oversight: instead of waiting for a scandal, it made scrutiny continuous, automatic and free for anyone to use.

Translating the reimbursement regulations into code made the rules themselves testable and let the project detect evasion patterns humans were unlikely to spot at scale.

Tweeting each suspicion to the accused politician made the deterrent public — accountability worked through visibility rather than prosecution.

Open-sourcing Rosie and building Jarbas turned a one-off campaign into reusable civic infrastructure that other countries could copy.

Crowdfunding the first 90 days and drawing on 600 volunteers proved that a serious oversight machine could be built by citizens, not just institutions.

When oversight is automatic and public, the deterrent is exposure itself: a bot that tweets every flagged claim changed politicians' behavior even where formal punishment rarely followed.

In 2018 the project was institutionalized under Open Knowledge Brasil, and founders moved into civic-tech roles — two joined the World Bank as fellows. The team turned the same data into Perfil Político, a platform for profiling candidates ahead of Brazil's 2018 elections. Its post-mortems were candid: congress could decline to investigate low-value claims (up to roughly US$20,000), and captchas on official portals slowed Rosie's scraping. Rosie still stands as a widely cited proof that AI can turn open data into citizen action — even when the watched institution resists.