Interactive demo · synthetic dataEvery record on this page is fabricated.

Ads Review Automation

A rules engine flags every pause. The model explains it, and may argue to keep an ad but never to pause one.

The production system behind this demo was engineered primarily by Prada Dipa — LinkedIn profile, opens in a new tab, and is maintained by Hady Satya. I managed and directed it — requirements, technical review, QA and rollout.

Read the case study

PROSPECTING CPA CAP60PROSPECTING ROAS FLOOR1.2RETENTION CPA CAP45VIDEO CPM CAP4.0GRACE DAYS7

HOLIDAY WINDOW

Product coverage is printed beside each pause, not used to withhold one.

8 PAUSE · 8 KEEP · 4 HELD (GRACE OR HOLIDAY)

SELECTED

MK-Americana-Video-01

SPEND: $340
PURCHASES: 0
ROAS: 0.0
CPA: —
CPM: $4.32
AGE: 11d

HOLIDAY OVERRIDE — ALL THREE MUST HOLD

In production the model applies this, from its prompt. Rebuilt here as a check.

✓ Holiday window open
✓ Overage within the allowance
✓ Ad name ties the product to the holiday

COMPUTED

KEEP

Kept: holiday window, CPM within 10% of cap, and the creative ties to the holiday.

ILLUSTRATION · NOT MODEL OUTPUT

Breach: CPM $4.32 vs $4.00 video cap — 8.00% over.
Recommended action: keep running through the holiday — re-evaluate after.
Reasoning: The holiday window is open, the 8.00% overage is inside the 10% allowance, and the creative is marked holiday-themed.

No model runs here. This is the three-bullet shape prompt v7 makes the model return for a flagged ad, filled by this page from the verdict above. In production the model may argue to keep a flagged ad, labelled as an override. It can never pause one the engine kept.

What is real, and what is simulated

Runs in your browser

Ported: rule construction from the JSON config shape, and all five comparison operators
Ported: STATUS_PRIORITY — the most severe status wins when several rules fire
Ported: the grace-period veto, including the conservative fallthrough when days_active is missing
Rebuilt, not ported: the three-condition holiday override. In production it is a rule in prompt v7 that the model applies, judging the holiday theme from the ad name. Here the theme is a flag set offline

The evaluator, the status priority and the grace period are ported from the production Python. 20 tests assert the behaviours it documents in prose, including the boundary day and the missing-days_active fallthrough.

Simulated

The Meta API pull, and the live model call
The model's own latitude: in production it may also keep a flagged ad on low spend or on a cited number that disagrees, labelled as an override. Nothing here reproduces that judgement
The model's note. No model runs on this page, and none wrote its text. Beside each ad the engine flags, the page fills the three-bullet shape prompt v7 requires (breach against its cap, action, one other number) from the computed verdict, labelled as an illustration. Ads the engine keeps get no note: in production the model never sees them
The thresholds are re-synthesised. The production caps are an unnamed employer's unit economics and are not reproduced — the architectural claim survives the substitution because it is a claim about shape, not about any number