Watch, create, publish, decide: the week is gone

An autopilot runs the loop, and decides what to scale or kill.

Demo on demo data: the autopilot reads the active campaigns, a campaign’s scale-or-kill thresholds, one line in the chat launches the agents, then the encrypted connections and the guardrails.
Autopilot panel on demo data: loop running, last tick, three active campaigns with hook, CTR, CPM, CPA, ROAS, results and spend.
Chat on demo data: after “launch the run”, the agent inspects the campaigns, launches the run and answers, creatives paused until validated.
A campaign’s autopilot strategy on demo data: scale and kill thresholds on ROAS and CPA, and the five-step decision funnel.
Settings on demo data: rules and thresholds the agent cannot bypass, hard budget cap per run, creatives paused, kill switch, tokens encrypted at rest.

The problem

Running Meta ads profitably is a loop of competitor watch, creatives, publishing and scale-or-kill decisions that eats a team’s week and still reacts late.

What I built

A desktop app where 22 isolated AI agents run a 9-phase pipeline, with an emergency kill switch checked at every phase. A genetic algorithm evolves the agents’ prompts against real campaign results.

Technical details

Fitness = 0.7·ROAS + 0.3·(1 − CPA), normalised
Winner/loser creative scorer on MiniLM text embeddings, with label leakage removed
Control plane bound to localhost only
Shell guard blocking exfiltration commands from scraped content

Stack

TypeScript + Bun
Tauri 2 (Rust)
Claude Agent SDK
SQLite
Meta Marketing API
Transformers.js (ONNX)

The result

296 tests (956 assertions) run in 5.9 s and cover the risky cases: injected instructions, forged requests, double clicks. Ad and store access tokens are encrypted at rest.

A similar project in mind?

Tell me about your business and when you need it. You get a proposal by email.

I want this system

More projects

SaaS platforms · COD stock & returnsYou do not know how much the courier owes youView the project