How I Ship YPH AI: ComfyUI, Cloudflare, and Cursor Agents (Founder Build Log)

TL;DR for busy founders

• YPH AI → Digital Content Logistics: winning format → Seed Video → Collab → Proof-of-Post at long-tail scale.
• Pain → The Sample Deficit: boxes scale linearly; affiliate posting and accountability do not.
• Edge stack → Vite + React on Cloudflare Workers, D1 for app data, Queues for render job dispatch.
• Render stack → ComfyUI on NVIDIA GPUs (autoscale pool); gateway on the GPU talks to ComfyUI locally.
• Dev loop → Cursor agents for iteration speed; founder review on positioning and compliance copy.

---

Why am I building YPH AI?

YPH AI (yourproducthere.ai) is The Digital Supply Chain for Social E-Commerce.

Operators on TikTok Shop and similar channels hit the Sample Deficit: every new affiliate traditionally means another physical box, another courier delay, and another creator who may never post. Logistics cost scales linearly. Creative output does not.

YPH does not claim to replace all physical sampling. VIP discovery and creative testing still benefit from real product in hand. YPH owns the long-tail scale phase once you already have a winning format (hook, pacing, demo, CTA that converts).

In product language that same object is the Seed Video. A Digital Sample bundles Seed Video + Collab (affiliate apply flow) + Proof-of-Post (delivery verification). That is Digital Content Logistics — not a digital download and not a substitute for testimonial claims affiliates did not earn.

What does the product loop look like for operators?

The operator workflow is deliberately boring in a good way:

1. Upload a Seed Video that already performs.
2. Open Collab so affiliates apply from a share link.
3. Affiliates receive a demonstration-style variant (not fake “I tried it” scripts).
4. Proof-of-Post unlocks rewards and future collabs.

Public copy uses winning format on landing and strategy pages. Seed Video stays in-app, legal, and ops surfaces. That split keeps affiliate-facing language clear while staying precise in Terms and admin review.

Compliance matters: YPH ships demonstration templates, not testimonial fiction. Scripts that imply first-hand product use without bona fide use are out of scope — see YPH Terms and seed review rubric on the product side.

How is the platform architected?

Think of two planes: edge control plane and GPU render plane.

Edge: Cloudflare + Vite + React

The public site and authenticated app share one Vite + React codebase. Cloudflare Workers serve marketing shells, the SPA under /app, and API routes.

D1 holds users, orgs, digital samples, collab state, and render job metadata. Cloudflare Queues decouple “user clicked generate” from “GPU is ready right now.” When capacity is tight, jobs stay QUEUED and dispatch retries — better than pretending renders are synchronous.

Marketing routes (/, /es, /experimental, /about, content articles) are prerendered where SEO matters. Collab share URLs get dynamic Open Graph via Worker HTMLRewriter so affiliates see the right preview when links spread.

Render: ComfyUI on NVIDIA GPUs

Video generation is not on Workers. A renderer gateway on each GPU machine talks to ComfyUI on localhost. Workflows handle face/voice alignment to the Seed Video while preserving product fidelity.

GPUs run on NVIDIA hardware with autoscale against queue depth — spin up when render-jobs backs up, scale down when idle. Stuck jobs get reaped and re-queued so operators do not stare at “RENDERING forever” after a VM dies mid-job.

That split is why quality benchmarks are public. Generic avatar tools do not prove seed-to-output fidelity for a specific SKU. Operators and TSP partners need to see occlusion, energy, and kinematics before they trust long-tail activation.

Where do Cursor agents help (and where they do not)?

I use Cursor agents the way I use any senior intern with infinite stamina: fast drafts, refactors, and glue code — with me as the editor who ships.

Agents helped with:

• React route and component iteration across marketing + /app surfaces.
• Worker API handlers, D1 queries, and queue consumer edge cases.
• SEO inject scripts, sitemap entries, and MDX content scaffolding (patterns I already proved on padron.sh).
• Test fixtures and observability log shapes when debugging render dispatch.

Agents did not replace:

• Category positioning and the hybrid physical/digital sample rule.
• Demo vs testimonial language in Collab output.
• Render quality thresholds and what we show on /experimental.
• Legal-safe claims in operator-facing copy.

If you are a solo founder reading padron.sh for Cursor vs Codex advice, this post is the adjacent chapter: same agent-assisted dev loop, different ICP (social commerce operators, not indie hackers picking IDEs).

What should you read next on YPH AI?

If you are an operator past MVP with a winning creative format:

• About YPH AI — category story and founder context.
• Quality benchmarks — seed-to-output proof.
• Long-tail affiliate activation guide — hybrid painkiller playbook.
• Ghosting after product sample — why Proof-of-Post exists.

If you are a builder curious about the stack: edge orchestration on Cloudflare, heavy ComfyUI on GPUs, queues in between — that is the shape. The product proof is in the pixels on /experimental, not in this blog post.