Built AI Image Data Pipelines for Model R&D
Cantina Labs
Built a results-first image scraping and validation pipeline to help Cantina Labs accelerate AI image model R&D.
Category / Type
AI systems engineering; R&D infrastructure; image data pipelines; dataset validation.
Client / Venture / Internal
Client work via Traverse in Space for Cantina Labs.
Context
Cantina Labs builds social AI characters that talk, perform, and interact with users. Lucas supported Cantina’s internal research team on image-generation R&D infrastructure and joined conversations with other image-generation teams in the industry, including Stability AI.
Wicked Problem
Cantina’s image research team needed a faster way to source, validate, classify, and expand image datasets for model development. The challenge was not just technical accuracy — it was speed. The existing approach had too much scaffolding and too few usable results: a lot of infrastructure thinking, but not enough validated images moving through the system.
The work required a balance that is easy to describe and hard to execute: move quickly enough to generate real data, but structure the pipeline well enough that the useful parts could later be scaled, repeated, and hardened.
Solution
Lucas designed and wrote code for an image scraping and validation pipeline that could crawl, qualify, deduplicate, label, and expand datasets according to configurable criteria. The system used RunPod-backed infrastructure and YAML-based configuration to move from ad hoc scraping toward a reusable dataset expansion engine.
The approach prioritized useful output before unnecessary architecture: prove the pipeline could produce valid datasets, then scale the parts worth hardening. The work touched technical areas such as phenotype representation, face labeling, deduplication, and image qualification, while keeping the core operating principle simple: results first, architecture around what works.
Outcome / Impact
Within weeks, the pipeline produced hundreds of thousands of valid images and established a reusable system for continued crawling, validation, and dataset growth.
Role / Contribution
AI Systems Engineer. Lucas was the solo Traverse in Space contributor on this work, working directly alongside Cantina Labs’ internal research team.
Supporting Proof / Artifacts
Reusable image scraping and validation pipeline.
RunPod-backed infrastructure for image validation and dataset expansion.
YAML configuration system for qualifying images against criteria.
Hundreds of thousands of valid images generated/qualified in weeks.