Client work
Early Childhood Copilot
Early Childhood Copilot is an AI assistant that helps Brazilian municipal managers look up the law, diagnose their municipality with public data and build a municipal plan.
- Technologies used
- Status
- Proof of concept
The problem
Anyone responsible for early childhood in a municipality, meaning children aged 0 to 6, has to decide with data scattered across many public databases and with law and rules spread across many places. Smaller municipalities often cannot hire specialists, data teams or consultancies. A dashboard of indicators shows numbers but does not help the manager understand, prioritize, plan, finance and monitor; the copilot covers that path, from question to plan.
What was built
Early Childhood Copilot is an assistant for municipal managers who do not need technical training. It brings together three layers that feed each other:
- a legal and guidance base, queried with plain-language questions, with the source cited in every answer;
- a municipal diagnosis, computed from 8 public data sources;
- a decision path in four steps: rights radar, prioritization of challenges, the municipal early childhood plan (PMPI) and goals.
The chat works in Portuguese and in English. The manager picks a municipality, and its profile enters the conversation so that answers start from local numbers. A feature counts as done only when it answers one question: which of the manager’s decisions does this support?
How it works
Law and guidance. The “Primeira Infância” collection in Libriq holds 22 documents: 12 legal norms, 4 guides and a 6-volume series on early childhood in the municipality. For each question, Libriq retrieves the most relevant passages and Claude writes the answer. The instruction to the model requires plain language, explained acronyms, a cited source and one concrete next step at the end. Before the laws are indexed, the system deletes the repealed wording, so that text no longer in force is not cited as current law.
Municipal data. The data passes through three layers: the raw data of each source, Parquet tables with 8 entities and validated contracts, and a curated PostgreSQL database with more than 5.7 million rows. The diagnosis gives historical series, trend and the municipality’s position among comparable municipalities. The code does the calculating; the language model only explains.
Decision path. This is deterministic too:
- the radar covers 8 rights and links each to an obligation, an indicator, a gap, a possible action and a citable legal basis, under the rule that missing data does not mean compliance;
- prioritization ranks challenges by 6 criteria, with versioned weights and 4 scenarios;
- the evidence library declares the grade of each intervention: robust, promising or good practice;
- the PMPI copilot keeps versions of the plan and feeds a goals panel.
Safeguards. By design, no personal data about children or families enters the system: the indicators are public and aggregated by municipality. The usage log is anonymous and stores no IP address, name or email. When the model refuses a question, the user gets a neutral message. The sampled citation audit has zero legal hallucination as its pass criterion.
Operation. The copilot runs on a dedicated AWS server with one-command deploy. The data pipelines are orchestrated with Airflow, and S3 storage is reached through the instance role, with no static key.
Who it is for and what stage it is at
The copilot is designed for municipal secretaries and technical staff, who would use it day to day, and for mayors, who need a short executive briefing to decide. The project adopts free access and priority for municipalities with low institutional capacity as design principles.
Early Childhood Copilot is a proof of concept running in production. Access needs a shared code, which suits a closed pilot and not an open product, and there are no individual accounts.
The weights, anchors and thresholds of the prioritization are provisional and are declared as such on screen until specialists validate them. The governance document on AI, data and ethics is still a draft awaiting specialist review. The plan calls for a pilot with 5 to 10 municipalities. One known limitation: citations reach the level of document and chapter, not of the article of law.
Where it connects to the rest of Sigil Labs’ work
The copilot is built on Libriq, which serves as the search layer over the collection of laws and guides. The municipal data can also be queried by agents: a read-only MCP server exposes five tools, three curated queries and two exploratory ones: one lists the Parquet tables and the other runs read-only SQL over them. In short, Libriq answers what the laws and guides say, and that server answers what the numbers are.
Frequently asked questions
What is Early Childhood Copilot?
Early Childhood Copilot, built by Sigil Labs for a client, is an AI assistant for managers of Brazilian municipalities who run policies for children aged 0 to 6. It answers questions about law and guidance with a cited source, builds a municipal diagnosis from 8 public data sources and walks the manager through to a municipal plan.
How does the copilot avoid inventing laws or articles?
Each answer is written from passages retrieved from a collection of 22 documents, and the instruction to the model forbids inventing a law or an article. When the collection does not cover a question, the copilot says it did not find the passage instead of quoting from memory. To check this, the project set up a sampled audit of 20 question-and-answer pairs, three of them deliberate traps.
Where do the numbers in the municipal diagnosis come from?
They come from 8 public sources, among them the IBGE censuses, the School Census, DATASUS health data, the CadÚnico social registry, SICONFI municipal finance data, vaccination coverage and SISVAN nutrition data. The numbers are computed by deterministic code, and the language model only explains them and never calculates them.
Does the copilot make decisions for the manager?
No, the copilot does not make decisions for the manager. The project's principle is that the AI supports decisions and never makes them, and the interface repeats that notice. Prioritization of challenges uses explicit, versioned criteria and weights, and the manager can switch scenario and watch the order change.
What stage is the copilot at, and who can use it?
The copilot is a proof of concept running in production. Access is by a shared access code, which suits a closed pilot and not an open product. The judgment parameters, such as weights and thresholds, are provisional until specialists validate them.
How does the copilot relate to Libriq?
Libriq is the copilot's corpus and search layer. It holds the laws and guides and returns the most relevant passages for each question. Answer generation sits in the copilot's own layer, among other reasons because it has to inject the municipal diagnosis into the prompt.