Ürünereklam

The problem

Ads in two languages, at the volume online shops need, usually mean an agency on a retainer. Even then nobody knows which version will work before the budget is spent.

How it is built

A Next.js app orchestrates durable Inngest jobs. A Python scraper reads the product page, with a hardened fetch path for difficult storefronts. Gemini writes the strategy, copy and frames, Veo or Seedance renders the video, ElevenLabs adds the voice, and Remotion adds word-timed captions, the watermark and the call to action. A separate studio produces static image ads and ad copy. Every variant then goes through a self-hosted model that predicts how people will react to it (TRIBE v2) on Modal GPUs, started and then polled, because a render outlives an edge function's timeout. Supabase holds the state behind row-level security with short-lived signed media URLs. Credits live in an append-only ledger with atomic charges and refunds, and checkout runs on Polar with prices in lira and dollars.

Decisions that mattered

• A failed step refunds itself

  The pipeline is made of durable jobs with a failure handler for each step. A step that dies refunds its credits exactly once and releases the run instead of leaving it stuck.

• Media is private by default

  Raw renders live in a private bucket, column-level SELECT is revoked and media links are short-lived signed URLs. No media address can be guessed, free tier included.

• Every ad is scored before anyone pays for it

  No ad leaves the studio without a score. The model ranks the variants before anyone puts a budget behind one.

Code from the repo

1async function safeRefund(
2  opts: Parameters<typeof refundCredits>[0],
3): Promise<void> {
4  try {
5    await refundCredits(opts);
6  } catch (err) {
7    console.error(
8      `[billing][RECONCILE-NEEDED] refund failed user=${opts.userId} ` +1
9        `amount=${opts.amount} ref=${opts.refType}:${opts.refId} ` +
10        `key=${opts.idempotencyKey ?? "-"}`,
11      err,
12    );
13  }
14}
15 
16// …
17await safeRefund({
18  supa,
19  userId: campaign.user_id,
20  amount: campaignCreditCost({
21    videoCount: brief.videoCount ?? 1,
22    videoModel: brief.videoModel ?? "seedance",
23    durationSec: brief.durationSec,
24  }),
25  refType: "campaign",
26  refId: campaignId,
27  idempotencyKey: `refund:campaign-all:${campaignId}`,2
28  description: "Kampanya başlatılamadı — otomatik kredi iadesi",3
29});

Where it is now

It grew from a hackathon demo into a bilingual product with GPU scoring. It is closed source now, and the paid launch is in preparation.