AdMagic

## Inspiration

Most ad budgets are spent finding out what doesn't work. Brands launch a set of creatives, pay for impressions, wait days for data, and only then cut the losers. Small teams and early-stage brands can't afford that kind of trial and error.

We kept asking one question: **what if you could know which ad will win before you spend a single dollar on it?**

That question became AdMagic.

## What it does

AdMagic predicts creative performance before launch. You upload your ad variations (images, video, copy) and AdMagic scores each one on how likely it is to drive clicks and return on ad spend.

At its core is **UNisal**, our proprietary scoring model. It breaks every creative down into the signals that matter:

- **Hook strength** \\( H \\): does the first moment stop the scroll?
- **Retention** \\( R \\): does attention hold or drop off?
- **CTA moment** \\( C \\): does the call to action land at the right time?
- **Overall drive** \\( D \\): a blended score that ranks creatives head to head

For any creative \\( c \\), the overall drive score is a weighted blend of the three signals:

$$
D(c) = w_h \, H(c) + w_r \, R(c) + w_c \, C(c), \qquad w_h + w_r + w_c = 1
$$

with weights tuned against real campaign outcomes. The result is a ranked shortlist with clear, explainable reasons, so marketers know what to launch, what to fix, and why.

We also built a **game layer** into the product. Users see three ads and pick which one they think performed best. The ground truth comes from real past campaigns, so every pick becomes a training signal that sharpens the model while teaching marketers to think like the data.

## How we built it

We built AdMagic on commercially licensed open-source multimodal backbones (**CLIP, SigLIP, Whisper** and **sentence-transformers**) to turn images, video frames, audio and copy into rich embeddings. On top of those sits our own fusion layer and prediction head, trained on real ad performance data from our beta users.

**Nebius AI Cloud** and **NVIDIA GPUs** did the heavy lifting. We used them to:

- Run multimodal embedding extraction across large batches of creatives in parallel
- Train and iterate on the UNisal fusion model fast enough to test multiple architectures within the hackathon window
- Serve low-latency inference so scores come back in seconds, not minutes

**SHAP-based explainability** sits on top, so every score comes with a reason a marketer can actually act on.

## Challenges we ran into

**Data honesty.** The easy path was to borrow research models and dress them up. Many of the best-known creative and attention models are released under non-commercial licenses, so we chose to build an original architecture from licensed components and our own benchmarks. Slower, but it's a model we fully own.

**Cold start.** A prediction model is only as good as its ground truth. With a lean beta, we had to be creative about getting labeled data, which is exactly why we built the three-ad game: it turns user engagement into training data.

**Video at scale.** Scoring video second by second is compute-heavy. Getting inference fast enough for a real product experience meant serious optimization of our GPU pipeline.

**Explainability.** A score alone doesn't help anyone. Translating model outputs into plain-language guidance a marketer trusts took more iteration than the model itself.

## Accomplishments that we're proud of

- Built **UNisal**, a fully original creative scoring model on our own benchmarks and baselines
- Running a live beta with **30 hand-picked users** and **100+ on the waitlist**
- Named a **top 10 finalist for CIBC x WEKH**
- Turned model training into a game that users genuinely enjoy playing
- Kept spend lean by not charging a single user before launch, focusing entirely on getting the model right

## What we learned

**Speed of iteration is everything in model development.** Access to serious GPU compute on Nebius turned week-long experiments into same-day ones, which completely changed how many ideas we could test.

**Marketers don't want a black box.** They want a second opinion they can argue with. Explainability isn't a feature on top of the model. It is the product.

**The best training data comes from users who are having fun.** The game outperformed every other data collection method we tried.

## What's next for AdMagic

- **Gate 1 validation:** hitting a 70%+ top-quartile CTR prediction rate across 500+ backtested creatives, where

$$
\text{Hit rate} = \frac{\text{creatives correctly predicted in the top CTR quartile}}{\text{creatives predicted in the top CTR quartile}} \geq 0.70
$$

- **Continuous learning loop:** feeding live campaign results back into UNisal so the model improves with every launch
- **Generative layer:** moving from "which ad wins" to "here's a better version," with AI-generated variations tuned to the score. Our first version is built; next comes more proof-of-concept testing to validate it at scale
- **Platform integrations:** plugging directly into Meta and Google ad accounts for one-click pre-launch scoring
- **Public launch in North America**, our beachhead market, backed by our upcoming pre-seed round

Our goal is simple: **stop treating ad budgets like casino chips.** AdMagic replaces gut instinct and hope-and-pray spending with prediction, so brands stop paying A/B tests to learn what doesn't work.