Henrique Brandao

I built and launched an LLM-powered conversational interface that lets security teams query risk posture data across 12+ million companies in natural language. I owned the back end fully, including Django and DRF APIs and Snowflake-backed retrieval to LangChain orchestration and Hugging Face and OpenAI model integration. I shipped the solution to AWS production in six weeks, with Kubernetes-deployed services, Helm-driven environments, and Terraform infrastructure. The system replaced rigid form-based queries with conversational discovery, reducing time to insight for customer security teams from minutes of clicking to a single prompt.

I guided a team of six engineers building a credit decisioning engine for Serasa Experian, Brazil's largest credit bureau, combining PySpark feature pipelines with neural-network scoring. The system grades creditworthiness for thousands of customer companies in production, processing high-volume application streams to detect fraud and assess risk. I owned architecture, sprint planning, and code quality standards across the engagement, and partnered with the data science team to productionize models.

I designed and shipped a compliance record migration pipeline at OneTrust that moved customer privacy and consent data across environments with zero data loss. Built on Django REST endpoints, the solution was deployed to Kubernetes via Helm and orchestrated with Terraform and Argo CD for fully GitOps-driven releases. I hardened the surface against partial-failure scenarios, idempotent reprocessing, and replay, which is critical for regulated workloads where rerunning a migration step can't violate consent semantics. I restored test coverage from near zero to more than 80% across three legacy services before any high-risk migration cut over.

I built a real-time vehicle license plate recognition pipeline at Multiway, running across hundreds of city cameras for public-safety operators. The solution combines OpenCV preprocessing with Keras and TensorFlow neural networks for character recognition and plate detection, optimized for sustained throughput against noisy live video streams. I tuned the model and the inference pipeline together: recognition latency, false-positive rate, and memory footprint were jointly constrained by hardware budgets at the camera-cluster level. I also conducted OWASP penetration tests on adjacent services and maintained the legacy Java Spring platform supporting the system.