The latest articles on DEV Community by Pritesh (@mannpriteshh). https://dev.to/mannpriteshh https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3947069%2F29857453-7967-4030-b5d0-c90f65fece80.jpg https://dev.to/mannpriteshh en Pritesh Sat, 23 May 2026 04:39:16 +0000 https://dev.to/mannpriteshh/i-built-an-ai-faceless-video-generator-in-2-months-heres-the-stack-266h https://dev.to/mannpriteshh/i-built-an-ai-faceless-video-generator-in-2-months-heres-the-stack-266h <p>Six months ago I started <strong>Keyvello</strong> (<a href="https://www.keyvello.com" rel="noopener noreferrer">keyvello.com</a>) — an AI video generator that turns a prompt into a complete short-form video in 2–5 minutes. Here's the technical breakdown for fellow builders.</p> <h2> The problem </h2> <p>Faceless creators on TikTok / YouTube Shorts / Reels spend 2–4 hours per video on scripting, voiceovers, B-roll, captions, and editing. Most burn out before they post 10 videos.</p> <h2> The stack </h2> <ul> <li> <strong>Frontend</strong>: Next.js 16, React 19, TypeScript, Tailwind CSS 4, Radix UI</li> <li> <strong>Backend</strong>: Next.js API Routes (App Router)</li> <li> <strong>DB</strong>: Supabase (Postgres + Auth + RLS)</li> <li> <strong>AI</strong>: GPT-5.5 for scripts, Fal.ai for images, ElevenLabs for voices</li> <li> <strong>Video</strong>: FFmpeg via <code>fluent-ffmpeg</code>, Sharp for image processing</li> <li> <strong>Storage</strong>: Cloudflare R2 (S3-compatible)</li> <li> <strong>Payments</strong>: Dodo Payments</li> <li> <strong>Compute</strong>: Vercel for the app, Modal for the video pipelines</li> <li> <strong>State</strong>: Zustand</li> </ul> <h2> The pipeline </h2> <p>prompt → GPT-4o script → scene splitter → parallel(Flux images + ElevenLabs audio) → FFmpeg composition (Modal) → R2 upload → status update</p> <h2> What surprised me </h2> <ol> <li> <strong>Modal beats running FFmpeg in Vercel.</strong> Cold starts on Vercel functions made 60s+ videos impossible. Modal webhooks solved it.</li> <li> <strong>RLS is non-negotiable from day one.</strong> Retro-fitting row-level security at 1K users is painful.</li> <li> <strong>Credit refunds need their own RPC.</strong> I hit a silent failure with <code>increment_user_credits</code> getting blocked by a trigger. Use <code>add_credits</code> instead.</li> <li> <strong>Users want templates, not raw control.</strong> I shipped a "blank canvas" mode early. Nobody used it. The 11 named templates (AI Stories, Fake Texts, Stick Animation, etc.) do 95% of generations.</li> </ol> <h2> What's next </h2> <p>Better lipsync for the talking-avatar templates. Tighter cost controls per template tier. Affiliate program.</p> <p>If you're building something in AI video, would love to compare notes — drop a comment.</p> ai nextjs saas showdev