Projects
Production systems, multi-agent AI architectures, machine learning implementations, and distributed backends.
Multi-agent orchestration for food-delivery post-order support (Reception on Claude Haiku + Order & Refund Specialists on Sonnet) with isolated MCP servers, deterministic refund rules in code, 4 tool guardrails, custom Portuguese BM25 + Snowball stemmer from scratch, and a 30-scenario τ-bench evaluation suite (8 adversarial, LLM-as-a-Judge, pass³).
User-Based Collaborative Filtering engine in production (UEFS). Compresses patient histories into 200-permutation MinHash signatures indexed in self-tuning LSH Forest prefix trees, with a Memory-First LRU cache (~200ms warm start), batch inference (300 patients/call), and 4-level graceful degradation.
Asynchronous telemetry ingestion and processing pipeline built with Elixir/OTP and Broadway, featuring configurable batching and backpressure, persistent background jobs via Oban, real-time Phoenix PubSub broadcasting, and Telemetry Metrics.
Multi-tenant backend API in NestJS and TypeScript with JWT authentication, role-based access control (RBAC), transactional concurrency control against race conditions, strict tenant isolation, and Jest/Supertest integration tests.
Concurrent distributed services in Go (goroutines, Gin, HTTP, and MQTT/Mosquitto) implementing the Two-Phase Commit (2PC) protocol (prepare, commit, abort) to coordinate distributed charging reservations across Docker containers.
Pure Python implementations of foundational ML models without black-box frameworks: Multilayer Perceptron (Backpropagation), Kohonen Self-Organizing Maps (SOM), Adaline (LMS Delta Rule), Rosenblatt Perceptron, Fuzzy Inference Controller, and Genetic Algorithms.


