Enter Form Logo
We're hiringPT
[ Mar 03, 2026 ][ Arthur Morais ][ 6 MIN ]

Reducing Entropy: How Enter's Engineering Processes Legal Chaos

With over 90 courts operating in a decentralized way, the challenge isn't a lack of rules — it's high technical variance. Here's how the Monitoring squad turns procedural noise into clean signal.

In Information Theory, Shannon Entropy1 quantifies the uncertainty or degree of "surprise" contained in a data source. Without a doubt, the Brazilian judicial ecosystem is an extremely high-entropy channel.

With over 90 courts operating in a decentralized manner, the challenge isn't a lack of rules — it's high technical variance: heterogeneous systems, inconsistent data standards, and signal instabilities that make capturing information a complex problem.

This post is the first in our series on how Enter's engineering squads are structured. Today we're opening the hood of the Monitoring team. Our technical role goes beyond moving bits; we act as a mechanism for entropy reduction at the source.

In this first stage of our product loop, we process the noisy signal from the external world and transform it into structured, predictable information.

The Scale of Input: 75 Cases Per Minute

To understand the squad's architecture, we need to look at the numbers. According to the Justiça em Números 2025 (CNJ)2 report, approximately 39.4 million new cases were filed in Brazilian courts in 2024.

That means the Brazilian judiciary publishes roughly 75 new cases every minute.

The Monitoring squad is responsible for capturing, classifying, and ingesting this continuous flow. But the complexity doesn't stop at public data. We operate in a hybrid model, also connecting to companies' "nervous systems" (otherwise known as ERPs).

To tame this complexity, we split our engineering into two distinct engines with specific design patterns: Public Discovery and Private Integrations.

1. Public Discovery: The Controller Strategy

The biggest risk for our clients is the "blind spot": a case that has been filed against them, but which they're not yet aware of. On the public data front, our challenge is discovery.

How do you monitor something that doesn't yet have a known ID in your database?

To solve this, we implemented a Controller Strategy. Think of the Controller as the orchestrator of a long-range radar. It doesn't perform the raw scanning itself, but intelligently manages who, when, and where we should search.

  • Complexity Abstraction: While we use strategic partners to handle the raw connectivity layer with the courts (resolving edge-case instabilities), the orchestration intelligence lives in our Controller.
  • Target Management: The system dynamically manages lists of "targets" (company tax IDs and legal names), deciding scanning frequency based on risk and relevance heuristics.
  • Noise Filtering: The Controller acts as the first entropy barrier. It ingests raw data, performs Deduplication (ensuring we're not processing the same event from different sources), and prepares the object for the next stage.

2. Private Integrations: The Provider Pattern

While the public side deals with volume and discovery, the private side (Client Integrations) deals with Connectivity and Protocol.

Large corporations have complex legacy technology ecosystems — from giants like Salesforce and Benner to proprietary on-premise legal systems. For Enter, these systems are rich sources of context (financial provisions, internal documents), but integrating them remains a classic software engineering challenge.

Here, we adopted the Provider Pattern.

We created a unified interface (ProviderProtocol) that abstracts the complexity of the external system. For Enter's platform core, it doesn't matter whether the data comes from a modern REST API, a legacy SOAP query, or a legacy database. The client-specific Provider is responsible for translating that local dialect into the platform's universal language.

The Future: Autonomous Integration Agents — Until recently, many of these connectors required constant maintenance of crawler scripts. We are now investing heavily in evolving these Providers into Autonomous AI Agents. Instead of rigid scripts that break whenever a client's ERP layout changes, we're training models capable of navigating, interpreting, and extracting data from these systems in a resilient way. (But that's a topic for a future post.)

The Big Challenge: Merge and Consistency

Where do these two worlds collide? At the Merge stage.

We frequently receive conflicting signals about the same case.

  • Source A (Court information from source X): Says the deadline is the 15th.
  • Source B (Client ERP): Says the deadline is the 12th.
  • Source C (Court information from source Y): Says the deadline is the 14th.

If we handled this naively, we'd end up with duplicate data or, worse, incorrect data. Our Data Consistency layer steps in here to create a Single Source of Truth.

We use reconciliation logic that employs language models to classify and deduplicate events captured from multiple sources. We validated the hypothesis that language models perform excellently at zero-shot learning, even in hard-to-control domains. However, tests conducted with internal datasets across several models revealed a minimal performance difference between large and small models. As a result, we were able to reduce costs by using smaller models while maintaining satisfactory performance. That said, we don't place blind trust in any model.

In this context, keeping humans in the loop is critical. Specialized validators ensure integrity in ambiguous cases, creating a feedback loop we use to continuously refine the hyperparameters used in our models for increasingly precise conflict resolution.

In his 1948 paper1, Shannon stated that "the fundamental problem of communication is that of reproducing at one point either exactly or approximately a message selected at another point". Furthermore, for Shannon, information is what reduces uncertainty. If I tell you something you already knew, the information is zero (because it didn't reduce your uncertainty). If I show you data that eliminates 50% of your doubts, I've given you 1 bit of information.

For the Monitoring squad, that "other point" is an ecosystem of 90 courts with millions of events and no obvious patterns. Our engineering exists to transform procedural noise into clean signal. By reducing the entropy of raw data, we ensure that, at the end of the line, our clients recover what Shannon defined as the essence of information: the freedom of choice grounded in certainty.

Being an engineer on the Monitoring squad means understanding that code is a means to reduce entropy in the real world. We'd rather deal with the computational cost of processing noisy data than assume the legal risk of ignoring a critical event. Without our hybrid ingestion layer, generative AI would just be a brilliant brain operating in the dark.

Footnotes

  1. Shannon, C. E. (1948). A Mathematical Theory of Communication. Bell System Technical Journal, 27(3), 379–423. https://ia803209.us.archive.org/27/items/bstj27-3-379/bstj27-3-379_text.pdf ↩ ↩2

  2. Conselho Nacional de Justiça. (2025). Justiça em Números 2025. https://www.cnj.jus.br/pesquisas-judiciarias/justica-em-numeros/ ↩