Legal Engineer — building reliable agentic systems for legal practice.
AI Engineer / Lawyer (PUC/SP). Former criminal law practitioner at top Brazilian firms, co-founder of Pindograma (data journalism — built Brazil's largest electoral poll aggregator).
Currently focused on modeling the legal domain with emphasis on steerability and explainability.
I believe good legal AI requires hybrid architectures: deterministic workflows for predictable, auditable paths + unconstrained loops for better "exploration/exploitation" during long-running tasks.
- 🔬 Innovation Resident @ InovaUSP — Building and validating GenAI products for legal practice using design thinking
- ⚖️ Integrating Brazilian court APIs (Datajud / BNP / MNI) into AI agents
- 🔓 Reverse engineering undocumented court systems via HTTP traffic analysis when official APIs don't exist
- 🔗 Designing multi-step tool-use patterns for nuanced legal reasoning during exploratory tasks
- 🧠 Working on neuro-symbolic approaches for high-stakes decisions (LLMs/SLMs for structured extraction + deterministic reasoning engine based on decision trees)
| Project | Description |
|---|---|
| research-squad | Multi-agent research system built with Effect + BAML. Hierarchical orchestration, structured concurrency, contract-driven TDD. Inspired by "How we built our multi-agent research system" by Anthropic. |
| TalentScore | Resume parser with deterministic scoring. LLM extracts structured data, then rule-based engine scores candidates - a promissing neuro-symbolic pattern for regulated industries. |
| harvest-mcp | MCP server that reverse engineers APIs from HAR files and generates TypeScript wrappers. LLM-powered dependency graph analysis. |
| inpi-agent | Minimal BAML agent with calculator + INPI database access. Reference implementation in Portuguese for the Brazilian dev community. |
Interests: effect systems, event-driven architecture, declarative DSLs, actor-based modeling, recursive agents with subtask spawning.
Happy to chat about similar problems, feel free to