Building an AI content operation that scales to millions of views

Executive overview

Running human creators is slow, inconsistent, and expensive — and you still don't know if the content will perform. Zuhair Lakhani replaced the entire creator workflow with AI-generated content posted via physical phone farms, generating 4.7 million views for one client in under four weeks.

The core insight is owning the full stack: content creation, account deployment, and performance feedback in one loop. Without all three layers, you have no data, and without data, the content never gets smarter.

The arbitrage is the feedback loop — content that learns from its own performance compounds faster than any human creator team.

From retail arbitrage to attention arbitrage

• Started with sneaker botting during COVID, then pivoted to retail botting (printers, pools, baby toys).
• Sneakers are illiquid; consumer goods have better throughput but no long-term brand.
• Applied Supreme-style hype tactics to a restaurant: waitlist drops every Wednesday, engineered scarcity, sold out for two months with 6,000 signups.
• Lesson: pulling a tactic from one industry into another creates temporary arbitrage before competitors catch on.
• Joined a TikTok creator program, realised he could run the same playbook for himself.
• First AI-content product: a poster sold on TikTok Shop using Midjourney + ElevenLabs, $100k revenue in month one.

Why human creator teams don't scale

• Standard model: hire college-age creators at $500–$1,000/month per 30 videos, manage via Slack.
• No accountability — creators post when they want, ignore guidelines, and still get paid.
• One full-time manager per 25 creators; ROI is negative once management overhead is counted.
• Content quality is unpredictable; you're optimising for one viral hit rather than a repeatable system.
• AI can recreate most of this content; the marginal cost of the human layer is hard to justify.

The phone farm: solving the deployment problem

• TikTok's device fingerprinting blocks Android emulators (Bluestacks no longer works).
• Solution: physical devices running custom software that swipes, watches, reposts, and comments to mimic human behaviour.
• Accounts are warmed up by searching specific terms, watching relevant content, and using an LLM to verify relevance before engaging.
• Warming algorithms are proprietary; three years of trial and error informed what TikTok actually tracks.
• Multiple accounts per device, each with a distinct persona (age, location, content niche) to enable A/B testing without too many variables.

The content system: format, generate, deploy

• Core format: slideshow with an AI-generated face on slide one, image-bank assets for remaining slides.
• Hook structure: name the pain point on slide one, offer increasingly impractical solutions, then insert your product as the easiest option.
• Synthetic faces are created by contracting someone's likeness, training a LoRA, then generating consistent images with Flux or Higgs Field.
• Templates are built from viral TikToks: paste a link, the platform reverse-engineers the format into a modular template.
• From that template, generate variants by changing hooks, swapping images, adjusting copy — all via prompts, not manual editing.
• Human review is kept for a final 5% check before publishing; AI is not yet reliable enough to skip this step.

Attention intelligence: the feedback loop

• Every post is tied to real account data — views, likes, bookmarks — from actual device deployments.
• Performance data feeds back into the content generation layer to improve future prompts and formats.
• Without owning deployment, there is no performance data; without performance data, content cannot improve.
• This closes the loop that manual creator workflows leave open: most teams never systematically feed results back into content strategy.

Double Speed platform demo highlights

• Carousel mode: select a template, auto-generate slide text, randomise image bank assets, queue to live devices in one flow.
• Hook-and-demo mode: upload app screen recordings, match with hooks, auto-generate overlay copy, render and queue.
• Account onboarding: input product description and target demographic; platform generates usernames, bios, and per-account search terms automatically.
• Image banks can be client-private or shared; clients scraping Pinterest often make them public.
• One client generated 4.7 million views across 15 accounts in under four weeks on the platform.

Business model and trajectory

• Agency phase: ran 15 brands simultaneously while in college, all AI-generated content, hit $40k/month.
• TikTok Shop: $100k revenue in month one; account banned when a 3PL manager was deported and 1,000 orders missed the two-day window.
• Raised $1M from a16z's Speed Run programme (their YC-equivalent accelerator).
• Thesis: the current "hire creators" cycle has run for 2–3 years; when every major brand does it, the format saturates and the next cycle begins.
• Mass-account strategy will persist as long as organic reach is follower-agnostic; only the winning formats will change.