AdSensei

Upload ads, get clear scores, one-line reasons, and precise fixes—then a ranked Top 10 by platform/audience in minutes. AdSensei helps designers iterate fast and marketers scale winners.

Comment

About the Project — AdSensei

Inspiration

We were done debating “which ad feels better.” AdSensei turns creative gut-feel into measurable signals so designers know what to tweak and marketers know what to scale.

What it does

AdSensei converts a folder of ad creatives into a clear, prioritized playbook:

  • Instant per-asset readout: tone & sentiment, product/category, objects/logos, OCR’d copy, color palette, composition notes, target audience, and best-fit platforms.
  • Actionable scoring across five core dimensions—Attention, Readability, Aesthetics, Brand Fit, Memorability—each with a one-line rationale so every number is defensible.
  • Smart recommendations: concrete fixes (CTA presence/placement, headline length bands, contrast/whitespace, logo size and anchor, modernized styling, eco-claims surfacing).
  • Campaign roll-ups: Top 10 with reasons, sentiment & tone maps, color trends, and Dimension Profiles (e.g., most readable vs. most memorable).

Example: single-image insight (condensed)

  • Scores: Catchiness 7, Aesthetics 6, Readability 8, Brand Fit 9, Memorability 7
  • Detected: logo “Rinso,” benefit copy (“WHITER THAN BRAND NEW”), detergent box, flags, dresses
  • Audience & platforms: 25–45, household care; strong for social, print, and nostalgia channels
  • Suggestions: add CTA, modernize layout, surface eco-benefit
  • Why: vibrant palette + crystal-clear benefit line; nostalgia boosts brand fit

Challenge alignment — high-value signals for recommendations

We extract diverse, high-value features that feed a recommender and generalize across campaigns:

  • Tone & sentiment → audience resonance and platform fit
  • Product/category → clustering and cold-start suggestions
  • Text/OCR & copy shape → headline length bands, readability, CTA presence/placement
  • Logos & brand anchors → recognition and consistency across variants
  • Objects & context → scene elements to inform targeting/use cases
  • Color & density → contrast/whitespace patterns tied to attention and legibility
  • Composition notes → focal order, safe-zone usage, rule-of-thirds hints
  • (Video-ready) AV signals → embeddings for distinctiveness/memorability

Why these help: they’re predictive proxies for clarity, fit, and distinctiveness—features a recommendation model can weight beyond historical CTR.

How we built it (high level)

We optimized for clarity, consistency, and speed—less black box, more “why”:

  1. Rubric first. Five dimensions that creative teams already speak: Attention, Readability, Aesthetics, Brand Fit, Memorability.
  2. Signals → explanations. Every score ships with a one-line rationale and a suggested change.
  3. Batch to decisions. Per-asset outputs aggregate into campaign-level Top-N, distributions, and Dimension Profiles.
  4. Sponsor tech in the loop.
  • creao.ai for a clean, skimmable UI.
  • Lava to route rationale prompts across models for sharper, less generic guidance.
  • Reka for robust vision-language grounding.
  • Inspired by AppLovin, we prioritized pre-spend insights (clarity, platform fit, iteration speed).

(The pipeline is model-agnostic and easy to swap as needs grow.)

Performance & parallel batch processing

Designed for fast (< 5 minutes) campaign reads and horizontal scale:

  • Parallel workers process assets independently (embarrassingly parallel).
  • Stateless jobs pull from a queue, emit compact feature vectors + rationales.
  • Reducer stage builds sentiment/tone maps, color trends, and Top 10 with reasons.
  • Caching reuses OCR/embedding results for near-duplicates.
  • Scale knob: increase worker count to hit tighter SLAs.

A simple, transparent scoring form keeps things explainable: [ S_{\text{composite}}=\sum_i w_i,z(d_i),\quad \sum_i w_i=1 ] where (d_i) are normalized signals per dimension and (w_i) are tunable weights.

Challenges we ran into

  • Subjectivity vs. consistency across wildly different aesthetics
  • No ground truth (early on) → relied on strong proxies before CTR/CVR labels
  • Signal fusion without letting a single metric (e.g., contrast) dominate
  • Edge cases: tiny text, busy backgrounds, logo-as-background patterns
  • Latency/cost while keeping batch UX snappy for 40–50 assets

Accomplishments we’re proud of

  • Explainable scorecards teams actually read—every number has a reason
  • Specific recommendations (e.g., “raise CTA ~120px,” “logo +18%”) instead of “make it pop”
  • Dimension Profiles that pick winners for a purpose (readable vs. memorable)
  • A clean campaign roll-up: sentiment/tone maps, color trends, and a demo-ready Top 10 with reasons

What we learned

  • Clarity beats cleverness—adoption rose when the rubric was transparent
  • Small, honest heuristics travel well: WCAG contrast, copy length bands, whitespace%
  • Rationales drive trust—teams change creatives faster when they see the “why”

What’s next for AdSensei

  • Performance linkage: learn (w_i) from CTR/CVR labels as they arrive
  • Platform presets: TikTok/IG/YT rules (safe zones, duration bands, subtitle norms)
  • Variant helper: headline trims, palette swaps, CTA placement experiments
  • Workflow & exports: reviewer notes, shareable one-pagers, and API hooks into creative pipelines

AdSensei: clarity, not guesswork.

Built With

Share this project:

Updates