Fiscal Digital
Timeline
From ideation to a pipeline in production. Complete transparency on what was built, how it works, what it costs, and where we're headed.
Step 0: the provocation
Brazil has more than 5,500 municipalities. Each one publishes an official gazette with contracts, procurement, appointments, and spending. That is millions of pages per year — public data, accessible by law, but impossible to follow at a human scale.
In 2016, the Serenata de Amor project demonstrated that algorithms can identify anomalous patterns in public spending. In 2019, Querido Diário digitized the official gazettes of hundreds of Brazilian municipalities. Two pieces of civic infrastructure built by civil society, together making something new possible.
If every gazette is already digitized, what prevents an algorithm from reading the same clauses an auditor would read — and doing so for every city, every day?
Fiscal Digital connects these two legacies: it uses the Querido Diário archive as its source and the Serenata principle as its method. The focus is the municipality — where everyday purchases happen, where street works are contracted, where political appointments are published.
We do not investigate. We do not accuse. We identify patterns that existing legal criteria define as irregular and make the result verifiable by any citizen. The primary source — the official gazette — is always one click away.
Timeline
- Done
Step 0 — Ideation
before April 2026
Discovery of Querido Diário → the provocation
- Done
Sprint 1 — Infrastructure Base
28 Apr 2026
Lambda, DynamoDB, SQS, Bedrock. Scheduled Collector for Querido Diário. Terraform with OIDC. 99 unit tests.
- Done
Sprint 2 — The First Agents
01 May 2026
5 specialized Fiscal Agents running in parallel: Procurement, Contracts, Suppliers, Personnel, Chief. Orchestration, caching, persistent memory. 129 tests.
- Done
Sprint 3 — Automatic Publishing
01 May 2026
Multi-channel (Reddit + X + RSS). Brand-gate against prohibited glossary. DRY_RUN active for validation before live.
- Done
Sprint 4 — Full pipeline in production
02 May 2026
End-to-end pipeline operating 7 days a week. To date: over 46,000 official gazettes read. AI layer via Amazon Bedrock for entity extraction from PDFs.
- Done
Sprint 5 — Site as primary channel
02 May 2026
Institutional site live at fiscaldigital.org. Public REST API (/alerts, /cities, /stats) and RSS feed. 50 cities covered, 10 specialized Fiscal Agents active. 180 alerts published after the Ciclo 4.1 calibration, every one linking to the original official gazette.
- Done
Sprint 6 — Infrastructure for scale
completed 09 May 2026
Every processed gazette leaves an auditable trail from the original PDF to the published alert. Operating cost lowered with targeted DynamoDB queries. Point-in-time recovery, single-command rollback, and risk-threshold calibration without redeploying.
- Active
Sprint 7 — Intelligence evolution
in progress
Supplier-level tracking (CNPJ): when the same vendor shows up in multiple findings, we flag the pattern. Public Fiscal evaluation set (golden set) measures precision before each release. Next: cover all 5,570 Brazilian municipalities as Querido Diário indexes them.
Current coverage
- 50 active cities across 22 states (top 50 Brazil by population + state capitals). 37 with alerts already published.
- Origin: Caxias do Sul, RS, municipal administration 2021 to present. 194 alerts published.
- Daily pipeline: 03:00 BRT, 7 days a week
- Next phase: gradual expansion as Querido Diário coverage grows
Costs and sustainability
Processing cost (live)
Full cost breakdown
The cumulative cost shown above includes AWS infrastructure since January 2026 and the fiscaldigital.org domain (GoDaddy). The complete service breakdown, USD→BRL conversion via PTAX BCB, and daily variation live on the FiscalCustos agent's canonical page.
See breakdown at /transparencia/custos →Funding model
Fiscal Digital is maintained by contributions from civil society. Every inflow and outflow is public. The code is open source under the MIT license.
Support on CatarseUnder evaluation
Public Fiscal evaluation set (golden set) with a target of 85% precision per agent before every release. Under study: Brazilian LLMs for data sovereignty, expansion to all 5,570 Brazilian municipalities.