Ad Detective

In progress

The live, in-browser demo for Ad Detective is being built. The full technical teardown is on the right →

Type a brand → watch an agent surface its copycats, with a vision model explaining each one.

Impact

image+text

multimodal

VLM

reasoned verdict

live

monitoring loop

Two modes off one engine: copycat detection and competitor research. A VLM compares creatives and explains the relationship, and turns thousands of unstructured ads into a queryable, structured dataset.

The problem

Winning ad creatives get cloned within days. Brands want to know who’s copying them; marketers want to know what’s working in a niche. Both hide in thousands of unstructured image-plus-copy creatives no one has time to review.

The approach

Reuse existing ad-library scraping infra (the hard part) and bolt multimodal AI on top. A VLM compares two creatives and explains the relationship, and extracts structured attributes. An agent ties it together; a monitoring loop watches for new clones.

Architecture

Brand name

one input

PROCESS

Ad-library scraper

creatives · copy · dates

MODEL

Multimodal embed

CLIP/SigLIP + text

STORE

Vector index

pgvector

MODEL

VLM verdict + extract

clone? + {hook, angle, offer}

OUT

Intel report + watchlist

ranked, explained

cross-cutting

LOOP

Monitoring loop

re-scrape · alert on clones

·Reuses existing scraping infra and bolts multimodal AI on top.
·A VLM both explains the copycat relationship and turns creatives into structured data.
·An agent runs the pipeline end-to-end; a scheduler watches for new clones.

How it was built

Phase 1

Fundamentals

✓Scrape a brand’s + a niche’s ads
✓Multimodal-embed (CLIP/SigLIP + text)
✓Your ad → nearest ads from other advertisers

Phase 2

Intelligence layer

✓VLM copycat verdict on pairs
✓VLM structured extraction
✓Copycat score = visual + copy + VLM

Phase 3

Agent + monitoring

✓Agent: brand → ads → copycats → report
✓Scheduled re-scrape, alert on new clones
✓Eval: detection precision/recall

Stack

CLIP / SigLIPa VLMpgvectoragent orchestrationschedulerad-library scraper

Framed for brand protection + legitimate competitor research over public ad-transparency data.

Source

Case study ↗Demo video

Want this on your team?

hello@miske.dev ← back to work