AI Video Generator A full-stack read of HUBX's paid-social funnel — ad creative through attribution, paywall assignment, model selection and unit economics — reconstructed from 1,586 decrypted requests off a single attributed install.

Every hop below was walked live and captured. The chain matters because each link carries a parameter forward — lose one and the paywall assignment changes.

The most revealing 15 minutes. Note what happens at 21:08:39 — before a single screen is shown.

The single most useful finding. The app does not hardcode a model — it asks the server which model to use, per feature, on every generation.

// the entire request body — no prompt, no model, no params
POST /process/img-video
{"imagePath":"user-inputs/<uid>/1786482866637"}
→ {"processId":"a5b5c8f4-…"}

// then poll every ~5s
GET /result/<processId>
→ {"status":"PROCESSING","remainingTime":4}   // counts 5→1, not %
→ {"status":"COMPLETED","videoUrl":"https://storage.googleapis.com/…"}

Adapty, public SDK key [PUBLIC_SDK_KEY_REDACTED]. Nineteen placements, each registered as its own named A/B test, and 34 distinct products. The SKU names are self-documenting.

Creative identity is carried all the way from the ad into the pricing engine. This is the part most teams never wire up.

// AppsFlyer returns the parked click to the fresh install
GET 1xnhtg.gcdsdk.appsflyersdk.com/install_data/v5.0/id6478868302
→ {
    "deep_link_value": "genericDance",     // ← the creative
    "af_status":       "Non-organic",
    "media_source":    "facebook_w2a",
    "campaign":        "facebook_w2a-campaign",
    "install_time":    "2026-08-11 21:09:09.120"
  }

// …and is immediately handed to the paywall engine
POST api.adaptytech.com/api/v1/sdk/attribution/profile/set/data/
{ "source": "appsflyer",
  "profile_id": "[ID_REDACTED]",
  "attribution_json": "{\"media_source\":\"facebook_w2a\", …,
                        \"click_time\":\"2026-08-11 21:04:18.840\"}" }

Click at 21:04:18, install at 21:09:09, attribution inside Adapty by 21:09:13. From then on every Adapty profile read carries media_source: facebook_w2a, which is what lets a placement resolve to a different price for ad traffic than for organic. ✓ verified — full payloads captured

The one that should raise an eyebrow: Microsoft Clarity is recording sessions inside the app. Combined with Adapty's placement data they can watch, screen by screen, where a given paywall variant loses people — not just that it converted worse.

• Resolve models server-side, per feature. A /features/{key}/model indirection costs almost nothing to build and buys the ability to swap providers, re-price, or A/B a model against a cheaper one without an App Store review cycle.
• Wire attribution into the paywall, not just into analytics. The AppsFlyer → Adapty handoff is the whole game: without it, price tests can't be read per creative, and creative tests can't be read per price.
• Test the intro and the renewal as a pair. Their SKU naming forces it. A $1.99→$9.99 ladder and a $6.99→$9.99 ladder land the same renewal at wildly different install economics.
• Countdown, don't percentage. Four-minute renders survive a 5→1 counter far better than a progress bar that appears to stall.
• Duplicate the winner, don't chase novelty. 162 of 194 ads are one creative running since April. The catalog is deep, but the buy is not diversified — the spend concentrates behind a single proven hook.