Darrell CalkResume
← All case studiesOVYRLORD

Founded and ran an AI media platform, concept to production

Over 78,000 lines of code: a social AI studio for image, video and audio, 181 versioned releases, and serverless GPUs that scale to 100 containers per model.

  • 78K+lines of code across the web app and GPU pipelines
  • 181versioned releases in nine months
  • 2,322LoRA styles in a curated library
  • 100GPU containers per model, scaled on demand
The OVYRLORD logo in neon pink and blue over a retro synthwave sun and grid

Role

Visionary, Chief Executive Officer & AI Engineer

The challenge

Generative image and video models run on expensive GPUs, while a consumer platform has to respond quickly, bill fairly and keep its AI systems shielded from public traffic. I founded OVYRLORD to put those models in users’ hands as a social studio, and built it from concept to production.

What I built

1. A social AI studio

  • A Next.js 16, React 19 and TypeScript web app with 28 pages, 116 API endpoints and 138 components.
  • Image, video and audio generation, including the Nano Banana image model, with a job queue that tracks every request from submission to delivery.
  • A social layer: public galleries, profiles, follows, likes, comments, direct messages, notifications, and achievements with experience points.
  • Sign-in with Google, Discord or an email link, plus one-time verification codes by email and SMS.
  • A blog, a public changelog, bug reports and support.

2. The generation pipeline

  • 82 ComfyUI workflows driving Python pipelines for request handling, workflow execution and result delivery.
  • Serverless GPU deployments on Modal across NVIDIA H200, H100 and L40S GPUs, running models such as LTX-2 for video, Qwen Image Edit and Z-Image. The video and image-editing models scale to as many as 100 containers each.
  • Photo and video generation and editing, up to 10 minutes long, using a curated library of 2,322 Low-Rank Adaptations (LoRAs): 1,265 for images and 1,057 for video.
  • Custom API middleware that protects backend AI systems, controls data handoff and isolates public traffic from core processing.
  • S3 storage for generated content and platform data.

3. Payments, data and operations

  • Credits and subscriptions on Stripe: credit purchases, monthly subscription credits, promotional credits, adjustments and per-model pricing.
  • An admin console with 22 admin endpoints for users, subscriptions, posts and the blog, including content controls such as forced-private posts.
  • A PostgreSQL data layer through Prisma, with 27 data models, migrated from MongoDB in December 2025.
  • Automated notifications and mass mailing for user communication and engagement.
  • Release management with automated versioning and changelogs, Git-based CI/CD, and production on a Hostinger VPS with an automatic A/B switch that alternates live versions to reduce downtime during releases.

By the numbers

  • About 78,500 lines of code: 69,500 in the web app and 9,000 in the Python GPU pipelines.
  • 1,092 commits and 181 versioned releases between July 2025 and April 2026, reaching version 4.3.
  • 116 API endpoints, 28 pages and 27 data models.
  • 82 generation workflows and a 2,322-LoRA style library.
  • System limits raised to 100 submissions per second, up to 1,000 per minute.