STAGE 1 / SCAN
Read every live ad in the category.
Active video ads only. The pet run read 220 on 15 Oct.
the ideaRead the ad library like a media buyer would, a whole category at once.
each dot here is an ad we have read · the glowing ones are brands that qualified
pipeline online
I engineered a
creative studio.
My own pipeline reads the ads in your category, finds the gaps, writes the hooks and makes AI-UGC to test them. Real creators then re-shoot the winners.
reading ads…
01 / hook swapper
One body. Three openings. Tap to swap the first 3 seconds.
Pick an opening and a persona: the hook swaps and the rest of the ad stays put. That's the whole business, three hooks on one body, tested.
SPEC AD · fictional brand
Spec ad for a fictional brand. Not a client.
What this tests
No performance numbers on this page. We haven't earned them yet, and we won't invent them.
02 / the pipeline, live
The machine behind the ad.
In its first runs, on 15 Oct, the pipeline read 520 ads from 219 brands across 5 categories, and wrote 54 hooks.
Seven stages, with real artefacts from the pet-supplement run. Brand names are hidden (k•••.com).
STAGE 2 / SCORE
Group by the brand that pays, then score.
62 brands, 34 with two or more ads, 18 qualified, 10 set aside.
the ideaAds are grouped by the site they link to, so a creator's page rolls up to the brand that pays them.
brandadsscoretier
STAGE 3 / TEARDOWN
Name the angle, the hook and the gap.
One real teardown of a top-tier brand, name removed.
the ideaThe same three questions for every brand, so the gaps compare across a category.
TEARDOWN · TIER A
- angle
- hook pattern
- format
- gap
STAGE 4 / HOOKS
Write three openings for the gap.
Each one is its own test of the same body.
the ideaHooks quote the brand's own claims and never invent new ones. A person edits every line.
STAGE 5 / GENERATE
Lock the persona, then shoot each take.
One reference still fixes the face, the room and the dog.
the ideaIf the persona drifts between takes, you're testing the actor, not the hook.
reference
take A
take B
take C
clip · 9:16 · 3 s
STAGE 6 / ASSEMBLE
Three hooks, one body.
Captions burned in. Only the first 3 seconds differ.
the ideaOne variable per test, and a hook swap moves a few hundred kilobytes, not a whole ad.
0s3s11s19s
V1
hook A
body
V2
hook B
body
V3
hook C
body
CC
captions burned in
STAGE 7 / OPERATE
Keep it moving, with a person in charge.
Schedules run the back office, on one server.
the ideaEverything up to the decision is automated. The decision isn't.
- 06:30New prospects reach the CRMMatched on the brand, never duplicated.
- 07:00Contacts are mergedOnly into empty fields.
- 18:30Approved emails become draftsNothing here can send.
- liveA reply raises an alertThe row updates itself.
A person sends every email.
1 / 7
03 / the studio
Three modules are live. More are being built.
live HOOK LAB
Pilot from $300
10 AI-UGC variants of one product: 3 to 5 hooks across 2 personas, plus a test plan that says what to run, at what budget, and when to stop each one. Five business days, one revision round.
live CREATOR COORDINATION
From $600/mo
4 to 8 human UGC videos a month from US creators. I handle sourcing, briefs with three written hooks each, shipping, deadlines and QA. You approve every creator and pay them directly.
live CREATIVE ENGINE
From $1,200/mo
Coordination plus 20 AI hook variants a month. The winning hooks get re-shot with real creators.
Retainers are month to month, with no contract. You own every video once it's paid for.
In the lab
- buildingWeekly creative reports from an Ads Manager export
- buildingA bench of creators matched to each brand
- buildingBrand brain: a living brief for every client
A chip changes to live only when it ships.
04 / the engineer
The engineering came first. The studio runs on it.
I'm a software engineer, B.Tech in Computer Science (AI), 2026. I design and build systems end to end, from the database schema to the cloud to the screen. The studio is one of them.
PujaBox
A full-stack platform I designed and built, with documented REST APIs and every decision on record.
- Next.js
- NestJS
- PostgreSQL
- AWS
Syna AI
A retrieval-augmented backend that answers from its own data, with pipelines that keep that data fresh.
- Python
- Vector DB
- LLMs
- AWS
Defect prediction
Models that flag defect-prone code modules, trained on NASA's MDP/PROMISE datasets, with a live dashboard.
- scikit-learn
- XGBoost
- Plotly Dash
This studio
The pipeline in the hero: ad intelligence, scoring, hook writing, AI-UGC and a back office. Designed, built and run by me.
- Python
- Automation
- One server
The studio is what happens when an engineer treats creative testing as a systems problem.
05 / the ask
Send your brand.
Three hook ideas for your current ads within 48 hours, written by a person. Free.
Rather talk it through? Book a 20-minute call.