AI Creative Technologist

Two hosts, one panel

• After Effects: six-tab panel (Tools, Translate, Subs, Voice, Music, Settings) with karaoke subtitle templates, per-platform safe-zone autofit, multi-size export and comp-level renaming.
• Premiere Pro: written from scratch on Premiere's DOM. Auto-captions (Whisper → SRT → import), CSV two-way translation, TTS dubbing straight onto the timeline, smart renaming, resize + safe-zone, QC pass.
• Shared core: panel UI, i18n, AI calls and key management are reused across both hosts. Only the ExtendScript host layer is rewritten per app.

What it does

• Translate: Gemini-powered, source language auto-detected, brand-name protection. Copy & Translate or in-place, plus CSV import/export.
• Subs: Whisper-timed captions from any audio, karaoke and static templates, stroke, highlight, BG box, font override, SRT/VTT import/export.
• Voice: ElevenLabs TTS with stability, similarity and speed controls; voice library, voice cloning and voice-change while keeping the original timing. Fit-VO-to-clip time-stretching.
• Music: trending track search across YouTube and TikTok, drag-drop into the project.
• Tools: multi-size export (1:1 / 4:5 / 16:9 / 9:16), packshot moves by keyword, project reorganization by language prefix, find & replace across selected comps.
• Settings: per-user API keys, 4 UI languages (EN, TR, PTBR, MK), 8 accent themes, first-run wizard with key validation, long-job retry + resume.

How it stays free and safe

• Zero-cost stack: Gemini for translation, Groq Whisper for ASR, ElevenLabs (per-user key) for TTS, yt-dlp for music. No proxy servers, no shared paid keys.
• Per-user keys: ElevenLabs is billed per user with their own key stored locally, so nobody can get banned for shared usage.
• Auto-updater: private GitHub repo behind a Cloudflare Worker with a fine-grained read-only PAT. Every installed panel checks once an hour, or on demand from Settings.
• Retry & resume: long transcriptions and dub jobs resume after a network drop instead of restarting from zero.

The six tabs

• Create: any model on fal.ai with that model's own controls, when you know exactly which model you want.
• Photo to Video: the main flow. Photo → first frame → talking video, side by side on one screen. Veo speaks the line itself, one call, no separate lipsync step.
• Scene: more than one person in the shot, with control over who speaks, who stays silent, and what each of them does.
• Motion Control: swap a different person into a video you already have. Same performance, rebuilt scene.
• Brief: paste a client brief and get back a full plan: cast, scenes, lines, b-roll. Edit it on a canvas, then render it.
• Assets: people, saved scenes, reference frames, b-roll, end cards, music. Ships stocked with hundreds of real first frames sorted by place.

How it stays honest

• Prompt sculptor: users pick dimensions (pose, outfit, lighting, camera, mood, delivery, accent, duration). The compiler stitches an English prompt and the text you read is the text that gets sent. Every category has a free-text field with an ✎ marker.
• Detached queue: submitting a render returns in a second. Close the tab, reload, lose wifi, kill the app, the work continues and the result still arrives. Verified by killing the server mid-render.
• Batch matrices: one photo × N scenes, or N frames × M lines, all through the same submit path. Re-running a batch you already paid for is free and instant.
• Story mode: hook → problem → solution → CTA joined into one ad. Clips are re-encoded (not stream-copied) so timestamps line up and audio never drifts.
• Cost control: every button shows an estimate before you click, a daily ceiling refuses anything that would cross it, and identical requests return the cached result instead of regenerating.
• Local by design: outputs land in real folders next to the app. Move the folder, your work moves with it. Nothing is synced, nothing leaves your machine except renders to fal.

Ship & install

• Zero-install bundle: unzip, right-click Open, paste your fal.ai key once. No Node, no npm, no dev account.
• Per-platform: Apple Silicon Mac, Intel Mac, Windows x64. The Node binary is embedded in the runtime folder.
• First launch downloads the reference frame library, then the app is fully offline for prompt work.

What ships in the platform

• Marketing site: hero, campaigns, teachers, programs, blog. Every block edited by the marketing team in Payload, changes reflect instantly.
• Booking engine: 1:1 lesson slots with teacher availability rules, no double-books, substitute options, credit and expiry handling.
• Checkout & billing: iyzico integration behind a provider-abstracted layer so EUR (Stripe) can slot in later without a rewrite.
• Student & teacher web apps: lesson dashboard, notes, recordings, progress, streaks and badges.
• Admin panel: content, operations, CRM and analytics in one place, on top of Payload with custom modules.
• Mobile app (Expo): single build, role-based (student mode / teacher mode), iOS + Android, ready to ship after web launch.

The hard parts

• Trilingual by design: TR / EN / DE, one source of truth in `packages/i18n`. UI copy, content models, emails and push, all localized.
• Two funnels, one CRM: the existing human setter / closer flow is preserved and instrumented, while a self-serve trial → checkout → onboarding path is A/B tested against it per lead source.
• Automation-first lifecycle: email + push for every lesson event, WhatsApp for sales and OTP only (deliberate cost decision). BullMQ worker for reminders, streaks, badges and sync.
• Zoom, Google Calendar, iyzico, WhatsApp Business: each integration behind its own adapter, so any one can be swapped without touching the booking engine.
• Extensible foundation: future modules (in-app video, EUR payments, e-book store, group events, AI speaking analysis) plug in without rework.

Stack

• Next.js 15 App Router · Payload CMS 3 · Expo (React Native)
• Turborepo · pnpm workspaces · TypeScript
• Supabase (Postgres) · Drizzle · tRPC · BullMQ worker
• Zoom · iyzico · WhatsApp Business API · Google Calendar

Where I use it

• AI automation & marketing video workflows: end-to-end pipelines that turn a brief into a finished, ready-to-ship marketing video with minimal manual work.
• Concept-to-creative pipelines: script → image → video → voice → subtitle in one chain.
• Localization at scale: copy, voice and visuals adapted per market with consistent brand.
• Internal tools & extensions: AI-augmented engineering. Full apps shipped end-to-end (this site, Marketing Studio, OptimiRyze).
• Workflow design: building reproducible prompt chains the rest of the team can use.

Tools in the stack

• Reasoning & code: Claude (Opus, Sonnet) · ChatGPT · Gemini
• Image: Nano Banana · ChatGPT Image · Midjourney · and others
• Video: Kling · Veo · Seedance · and others
• Voice & audio: ElevenLabs (TTS, voice clone) · OpenAI Whisper (ASR)
• Inference platforms: fal.ai · Replicate
• Production tools I built using AI: Marketing Studio · OptimiRyze · this site