OpenArt AI Platform Work

Problem

A fast-moving AIGC product needs to integrate many model providers, ship new generation workflows, organize user assets, instrument behavior consistently, and migrate data safely without interrupting production traffic.

Solution

Contribute across the AI product/platform layer: integrate model-backed workflows, own Projects & Folders across frontend and data model boundaries, build Modal/FastAPI services, and execute a large Firestore migration with checkpointing, throttling, sample verification, rollback support, and production-safe rollout practices.

Architecture

• Model workflow layer integrates 20+ generative AI models across about 12 vendors for text-to-image, image/video generation, video editing, reference-to-video, chat-to-edit, and 3D-world capabilities.
• Projects & Folders spans Firestore data modeling, indexes, URL-canonical state, frontend chrome, analytics taxonomy, and LaunchDarkly rollout.
• Modal/FastAPI microservices support workload-specific services such as thumbnail generation, LTX 2 video, and feedback intelligence.
• Firestore migration uses checkpointed range scanners, live throttle control, orphan-data skipping, verification samples, rollback support, and throughput monitoring.

Engineering Decisions

• Treat model integration as product infrastructure, not one-off API calls: workflows must be observable, rollable, and maintainable across providers.
• Use typed analytics registry patterns to reduce scattered event names and property drift across product surfaces.
• Design migration tooling around recoverability and live-traffic safety instead of a single long-running cursor.
• Keep feature rollout and production polish in scope, because AI product quality depends on frontend state, asset organization, and operational confidence.

Outcomes

• 20+ model integrations contributed to about 10.7M generations in 90 days, around 24% of platform volume.
• Projects & Folders organized 250K+ generation records while partially flag-gated.
• Thumbnail service handled about 437K calls/day, around 13M/month.
• Firestore migration processed about 922M documents with peak throughput around 3,665 docs/sec.

Next Steps

• Keep this as a public-safe case study without exposing proprietary code, internal diagrams, or confidential implementation details.
• Use this case study to anchor the portfolio narrative before smaller public repos like SnapLink and GalaGate.
• Add sanitized architecture diagrams or screenshots only if they can be created without revealing company-private information.