Work

Selected public and enterprise work.

A curated list — not exhaustive. Enterprise items are presented without confidential details, focused on architecture, constraints, and impact.

Featured

Open source · Applied AI

CorridorKey-Runtime

Native runtime and OFX plugin for neural green screen keying in DaVinci Resolve, with GPU acceleration for NVIDIA RTX and Apple Silicon. Built in collaboration with Corridor Digital, with a deterministic inference path designed for real post-production workflows.

Role: Engineering lead — runtime architecture, OFX plugin, cross-platform inference (CUDA / MLX), packaging.
Impact: Brings Corridor's keying model into a professional editing tool as a production-ready plugin, without a cloud dependency.
Stack: C++ · Python · CUDA · MLX · ONNX · OFX

Other selected work

Enterprise · Computer vision platform

SODA Vision — AB InBev

End-to-end computer vision platform used across AB InBev factories. I led architecture for the ecosystem, separating a cloud-based editor and management layer from a deterministic runtime. Introduced a domain-specific language to standardize industrial vision pipelines and a standardized inference node architecture with offline execution and industrial data publishing.

Role: Software Architecture — platform design, DSL, runtime, inference node architecture.
Impact: Standardized how computer vision is built, deployed, and operated across a global manufacturing footprint. Co-created the LAHBS Box initiative, which evolved into a commercial offering.
Stack: Python · C++ · PyTorch · OpenCV · Docker · Linux

Enterprise · Applied AI

Visual inspection system — Colgate-Palmolive

A deep-learning visual inspection system that replaced legacy sensors on high-speed production lines across facilities in the USA and Mexico. Delivered through Baunilha Systems: calibration workflows, GPU-accelerated industrial PCs, on-site validation and production rollout.

Role: Architect and tech lead — model integration, deployment, validation, on-site rollout support.
Impact: Reduced false rejection rates and enabled real-time quality analytics on high-speed lines.
Stack: Python · PyTorch · OpenCV · Industrial PCs · Linux

Enterprise · Systems

Marvitech — Multi-tenant backend & device fleet

A multi-tenant backend in .NET and PostgreSQL, plus fleet management for Raspberry Pi devices with reliable OTA updates in low-connectivity environments. Paired with an offline-first Flutter mobile app with background synchronization.

Role: Architecture and delivery — backend, device fleet, mobile sync.
Stack: .NET · PostgreSQL · Flutter · Raspberry Pi · Linux

Open source · Applied AI

vlm-ocr

OCR experiments with multimodal models and local inference workflows.

Role: Author.

GitHub

Open source · Developer tools

macos-blueprint / blueprint

Reproducible development environment tooling for macOS and Fedora Atomic / Bluefin. Small, opinionated tools for keeping engineer machines consistent.

Role: Author.

Open source · Computer vision

face-access / severinopi

Face recognition access control projects for PC and Raspberry Pi — practical examples of deploying computer vision on constrained hardware.

Role: Author.

Experience

Roles

Present

Baunilha Systems

Principal AI Systems Architect & Tech Lead

Leading consulting engagements focused on applied AI, computer vision, and software systems for real-world operations.

Colgate-Palmolive — architected and deployed a deep-learning visual inspection system replacing legacy sensors across production lines in the USA and Mexico. Led validation and on-site rollout.
Marvitech — multi-tenant .NET / PostgreSQL backend, Raspberry Pi fleet management with OTA in low-connectivity environments, and an offline-first Flutter app with background sync.

AB InBev / Ambev Tech

Software Architecture

Led the engineering of SODA Vision, AB InBev's proprietary end-to-end computer vision platform used across factories worldwide.

Designed the core architecture of the ecosystem, separating cloud-based editor and management workflows from a deterministic runtime.
Created a domain-specific language to standardize industrial vision pipelines.
Developed a standardized inference node architecture with offline execution and industrial data publishing.
Co-created the LAHBS Box initiative, which evolved from an intrapreneurship program into a commercial offering.

Foundation

Industrial systems foundation

Before focusing more heavily on AI systems and computer vision, a substantial part of my early engineering work was in industrial software and mission-critical environments. These projects formed the systems thinking that runs through everything I build today.

Airport infrastructure · Safety-critical systems

Emergency systems software — Infraero

Embedded and control software for airport fire trucks deployed across Brazilian airports. Built for operational reliability in safety-critical, time-sensitive environments where system failure has direct consequences.

Role: Software development — embedded control systems, operational interfaces, field deployment.
Impact: Deployed across multiple Brazilian airports in a high-consequence operational context.

Offshore infrastructure · Mission-critical systems

Control systems software — Petrobras

Software for control and monitoring systems in offshore oil infrastructure. Designed around high-consequence constraints where reliability, fault tolerance, and deterministic behavior are non-negotiable.

Role: Software development — control systems, monitoring, industrial protocols.
Impact: Deployed in offshore environments where system availability is directly tied to operational and safety continuity.

Public utility · Industrial control

Infrastructure control systems — AySA

Control software for sluice gate and sanitation infrastructure in Buenos Aires' water and wastewater network. Public utility-scale systems with continuous operational requirements and no tolerance for downtime.

Role: Software development — control systems, infrastructure software, PLC integration.
Impact: Contributed to the reliable operation of public sanitation infrastructure serving the Buenos Aires metropolitan area.

Agricultural genetics · Technical automation

Automation and systems software — TMG

Software systems for an agricultural genetics and crop improvement operation. Spanned automation, data management, and domain-specific tooling for precision agricultural workflows.

Role: Software development — automation systems, operational tooling, process integration.
Impact: Brought engineering discipline to domain-specific operational workflows, reducing manual overhead in specialized genetics processes.