Metametrics translated computer vision research into a real manufacturing workflow, using live defect detection and process monitoring to make quality control practical on commodity hardware.
Executive summary
Computer vision for industrial quality control
Metametrics was a startup focused on computer vision systems for industrial quality control, developed as part of CESAR School's startup acceleration program. As tech lead, Vinícius oversaw the design and implementation of process monitoring systems in collaboration with computer science students, bridging applied machine learning research with real manufacturing constraints.
Context
Why the project existed
The project emerged from CESAR School's entrepreneurial ecosystem, which pairs student teams with real-world industry challenges. Metametrics tackled a common but costly problem in manufacturing: automated, real-time defect detection and process quality monitoring without requiring specialized hardware beyond standard industrial cameras.
Problem
What the system needed to solve
- Manual visual inspection — slow, expensive, and inconsistent across shifts.
- Rule-based machine vision systems — brittle to lighting changes, part variation, and new product lines.
- Costly proprietary inspection equipment — inaccessible to smaller manufacturers.
The goal was to build a software-first solution that could run on commodity hardware, adapt to new inspection tasks with minimal retraining, and integrate with existing production lines.
Technical approach
Detection, monitoring, and deployment
Detection pipeline
YOLO-based object detection balanced inference speed with accuracy for real-time defect localization and classification.
Process monitoring
Part throughput, cycle time variance, and anomaly rates turned raw detections into production intelligence.
Deployment
FastAPI inside Docker containers kept inference close to the production line with minimal latency.
Leadership & collaboration
What the role involved
- Setting the technical direction and architecture decisions for the CV pipeline.
- Mentoring a team of CESAR School students on model development, evaluation, and deployment practices.
- Translating customer requirements into engineering specifications.
- Coordinating the integration between ML model outputs and the production monitoring dashboard.
This role was the first direct experience managing a technical team and navigating the tension between research-quality models and production-grade reliability.
Outcome
Real-world quality control data is noisier and more varied than benchmark datasets — domain adaptation matters more than model architecture in this context.
Fast iteration loops and close collaboration with the end-user (production floor operators) are as important as model performance metrics.
The startup experience sharpened an understanding of product thinking alongside engineering depth.