Projects

First Brazilian open-source project incubated by Apache Foundation. An MLOps platform that helps data science teams ship production ML with a high-scale, low-latency, language-agnostic architecture. Simplifies the path from exploration to deployment.

• Apache Foundation incubator project, the first from Brazil
• PwC Innovation Value Award winner (2018), Retail category
• Speaker at Apache conferences, QCon, PAPIs, SEI Saturn
• ~413 combined GitHub stars across Marvin ecosystem, 33 forks, 585 commits
• Published paper: 'Marvin — Open Source Artificial Intelligence Platform' (PAPIs 2017)

End-to-end streaming architectures for 50+ enterprise customers with AI/ML integrations. Designed architecture patterns for ML pipelines, RAG, Multi-Agent Systems, and agentic workflows on Confluent's DSP stack. Led technical PoCs for high-profile LATAM accounts.

• Architected solutions for 50+ enterprise customers, supporting $15M+ in deals
• Led AI/ML PoCs: anomaly detection, agentic workflows, MCP, A2A integrations
• Authored 'CrewAI Crews as Apache Flink Jobs' and 'Flink Multi-Agent System Design'
• Contributed to Confluent MCP, Flink Agents (FLIP-531), Streaming Agents, Lightning Tables
• Technical architect for Blip, iFood, Pipefy, Stix, Michelin, Superlogica

First-of-its-kind fraud detection platform for the dental insurance industry, processing terabytes of dental images using AI to identify fraud patterns. Built from zero to production.

• Built from scratch (zero to one) as co-founder
• Identified ~0.5% fraud rate in its first release
• Processed terabytes of dental images using AI
• Enabled dental payers to take corrective and educational actions
• Globally distributed team across USA, Brazil, and Philippines

Scaled and optimized Overjet's dental AI inference pipeline, drastically improving latency and reducing infrastructure costs. Also conceptualized and built an innovative fraud detection MVP in one quarter.

• Improved RTA latency of AI inference pipeline by ~20x
• Reduced unit costs by ~6x through ML component rewrites and K8s optimization
• Processed millions of images daily across hundreds of production services
• Built fraud detection MVP (first of its kind) in one quarter
• Achieved 99.9% platform uptime with 90% reduction in incidents

First SaaS platform in Latin America for e-commerce data collection and product classification. A crawler platform for structuring and matching products at scale, handling price monitoring and product deduplication across major e-commerce platforms.

• First SaaS platform of its kind in Latin America
• Team of ~20 people, built in 6 months
• Spawned two derivative products: Boomage and Booska
• Price monitoring and product deduplication at massive scale
• Combined NLP and computer vision for product structuring

Deep learning system for visual product recognition using computer vision. A scalable, low-latency API for white-labeled image-based product search across e-commerce platforms, handling millions of classified images with sub-5-second response times.

• First stable version delivered in under 4 months
• ML at scale: millions of classified images with 5-second response requirement
• Distributed KNN for visual similarity search
• GPU-accelerated deep learning pipeline
• White-labeled API for e-commerce integration

Web product search and price comparison platform. Architected as a web scraping system that independently compares competitor prices across multiple e-commerce sites, structuring large-scale web-derived data.

• Automated price comparison across multiple e-commerce seekers
• Web scraping at scale with NLP-based data structuring
• Real-time competitor price monitoring
• Beta version deployed and validated

AI-driven system that assists software architects using case-based reasoning and architectural tactics. Suggests solutions based on non-functional requirements analysis, reusing proven architectural patterns.

• AI-powered architectural decision support
• Case-based reasoning for solution reuse
• Quality attribute analysis and non-functional requirements mapping
• Part of M.S. thesis research at IPT

DRM-protected digital book distribution ecosystem supporting EPUB and PDF formats. An economical alternative to Apple and Adobe solutions for small publishers, with apps across iOS, Android, and web.

• Beta launched in 6 months with minimal team
• Cost-effective alternative to Apple/Adobe DRM solutions
• Full ecosystem: server, iOS app, Android app, web reader
• Custom encryption and DRM implementation

Data & ML pipeline blueprint for legal domain projects with Python SDK encapsulating ABBYY FineReader Engine for OCR solutions, making C++ capabilities accessible to ML scientists.

• Reduced ML scientist prototyping time/complexity significantly
• Python SDK wrapper for C++ ABBYY FineReader Engine
• Used across PwC Labs legal domain projects
• Managed globally distributed team (Argentina, India, USA)

MCP Server implementation for Trello board management, demonstrating how Model Context Protocol can be used for agentic tool integration. Featured at Austin Python community meetup.

• Featured at Austin Python meetup
• Demonstrates MCP server architecture and agentic patterns
• Community education contribution on emerging AI protocols