AdsPilot AI — AI Ads Optimization Platform Blueprint

AdsPilot AI — Build an AI That Runs Your Ads

From Ad Creation to Optimization — One AI System Does It All

A full blueprint for building an AI-powered ad management platform. Read this, and you'll understand the entire vision — from architecture to pricing.

Last updated: 2026-03-21

Ad platforms supported

AI agents working together

100+

Ad variations tested automatically

$28

Starting price vs $550+ agencies

What's actually wrong with how businesses run ads today?

Here's the reality: businesses pay ad agencies $550 to $2,800+ per month, and they still end up staring at confusing reports themselves. They don't know which ads are working, which ones to kill, and where the budget is leaking. AdsPilot AI fixes all of that.

Quick Summary

AdsPilot AI is a 6-layer automated ad management system — from content creation to budget optimization
9 AI agents act as the system's brain, working alongside an AUTOPILOT engine
Built in 6 phases over 30 days, with clear pricing plans for every business size
Designed for businesses that want to cut ad costs without cutting performance

👔

Sam — SME Owner

Ad Spend $1,400-5,600/mo

No time to monitor ads, agencies are expensive, no idea where the money goes

📊

Maya — Marketing Manager

Ad Spend $5,600-28,000/mo

Can't keep up with multiple platforms, manual optimization wastes time and leads to mistakes

🛒

Alex — E-commerce Owner

Ad Spend $840-4,200/mo

ROAS keeps fluctuating, doesn't know how to adjust ads, reports make no sense

The ad budget that leaks every month adds up to tens of thousands a year — AI can fix that starting at $28/month.

What exactly is AdsPilot AI?

An AI system that creates, tests, and optimizes ads automatically — end to end, in one place. No agency fees. No manual optimization. The AI handles everything from writing ad copy to reallocating budget across platforms.

📊

AI Ads Analyst

Analyzes ad performance across every platform in real-time, with alerts when something's off

✍️

AI Creative Generator

Writes ad copy automatically — 6 headline styles, localized for your market

🤖

AUTOPILOT

Creates and tests 100+ variations using Thompson Sampling — the heart of the system

📱

Multi-Platform

Google, Meta, TikTok, LINE, Shopee, Lazada — every platform in a single dashboard

🧠

Multi-AI

Claude, GPT-4o, Gemini via OpenRouter + Direct API — picks the best model for each task

🌍

Market-Aware

Built with local market context from day one — not a generic tool with a translation layer

What makes it different from everything else?

🌍

MARKET-AWARE

🤖

TRUE AI (MAB)

🔄

END-TO-END

📱

6 PLATFORMS

💰

FROM $28/MO

Dubai Skyline at Sunset — visual break before architecture deep dive

Take a breath. The architecture section is next.

How does the 6-layer architecture actually work?

The system is split into 6 layers, each with a clear job. Understand this, and you'll see the full picture.

🌐 Presentation Layer

Next.js 15 (App Router) + PWA + Slack Bot + LINE Bot | TailwindCSS + Shadcn/UI + Recharts + Zustand

🚪 API Gateway Layer

📦 Application Layer

🤖 AI Orchestration Layer

Temporal Workflows | 9 AI Agents | AUTOPILOT Engine (Picasso + Darwin + Sherlock) | LLM Router

🔌 Connector Layer

🗄️ Data Layer

PostgreSQL 16 (Core) | TimescaleDB (Metrics) | Redis Stack 7 (Cache) | Qdrant (Vector) | NATS JetStream (Events)

6 layers, each with a clear responsibility — a team member can own one layer without needing to understand the whole system.

See the full tech stack (complete table)

Layer	Technology	Why This Choice
Monorepo	Turborepo + pnpm	Fast builds, shared packages
Frontend	Next.js 15 (App Router)	SSR, RSC, TypedRoutes
Styling	TailwindCSS + Shadcn/UI	Utility-first, accessible
State	Zustand + TanStack Query	Lightweight + server cache
Backend	Go 1.22 + Fiber v3	High performance, type safety
DB Access	SQLC	Type-safe SQL to Go code gen
Auth	JWT (HS256) + bcrypt(12)	Stateless, secure
Core DB	PostgreSQL 16	ACID, JSONB, extensions
Metrics DB	TimescaleDB	Time-series optimization
Cache	Redis Stack 7	Session, rate limit, pub/sub
Vector DB	Qdrant	Semantic search for creatives
Queue	NATS JetStream	Lightweight, at-least-once
Workflow	Temporal	Durable async workflows
AI Router	OpenRouter + Anthropic	Multi-model, cost optimization
Container	Docker + k3s	Lightweight Kubernetes
CDN/WAF	Cloudflare (Free)	DDoS, SSL, caching
CI/CD	GitHub Actions	Automated pipeline

How do the 9 AI agents work as the system's brain?

Nine AI agents work together, each using the AI model that's best for its job. Commander is the team lead — it takes your request, breaks it into tasks, and assigns them to other agents running in parallel.

🎖️

Commander

Claude Opus — task delegation

📊

Analyst

GPT-4o — data analysis

✍️

Creative

Claude Sonnet — ad copy

👥

Audience

GPT-4o — targeting

💰

Budget

GPT-4o — budget allocation

🔮

Forecast

Gemini Pro — predictions

🕵️

Competitor

Gemini Pro — competitor analysis

📋

Reporter

Claude Sonnet — summaries

🔍

Scout

GPT-4o Mini — trend hunter

Why use multiple AI models? Each model has different strengths — Claude writes excellent copy, GPT-4o is precise with numbers, Gemini has a massive context window perfect for research. The system picks the best model for each task instead of forcing one model to do everything.

Cursor Prompt: "Build me an AI agent router that picks the best LLM for each task type — Claude for writing, GPT-4o for analysis, Gemini for research"

How does AUTOPILOT work as the core engine?

This is the feature that makes AdsPilot different from every other tool. It creates 100 ad variations, tests them, and optimizes automatically using Thompson Sampling — not simple rule-based logic.

🎨 GENERATE

Picasso creates 50-100 ads

→

🚀 DEPLOY

Push to live platforms

→

📚 LEARN

Collect data for 3-14 days

→

⚡ OPTIMIZE

Darwin allocates budget

→

📊 REPORT

Sherlock summarizes results

Thompson Sampling — simpler than it sounds

Think of it like 100 slot machines (= 100 ads). You need to figure out which ones pay best. Thompson Sampling does that by:

Ad #1

ROAS 3.2 🟢 Scale Up +30% Ad #2

ROAS 2.1 🔵 Maintain Ad #4

ROAS 1.2 🟡 Scale Down Ad #5

ROAS 0.5 🔴 Pause

Winners get more budget. Losers get cut. All automatic — no manual optimization needed.

Safety constraints — so the AI can't blow your budget

✅ Min test period: 3-14 days before optimizing

✅ Min impressions: 1,000 before making decisions

✅ Max budget per ad: 30% of total

✅ Human approval when budget change > 50%

✅ Confidence ≥ 0.8 before scaling up

✅ Can't pause ALL campaigns at once

The AI makes decisions on its own, but safety rails prevent disasters — big budget swings need human approval, and it can't shut everything off at once.

See the score calculation + decision tiers

Score = (ROAS/10 x 0.4) + (CVR/0.2 x 0.3) + (CTR/0.1 x 0.2) + (Confidence x 0.1)


    🟢 SCALE UP   ≥80th percentile → increase budget +20-50%

    🔵 MAINTAIN   40-80th percentile → keep testing

    🟡 SCALE DOWN 20-40th percentile → reduce budget -20-30%

    🔴 PAUSE     <20th percentile → stop this ad

See the data flow — how information moves through the system (4 main flows)

Flow 1: Ad Data Sync (every 1-6 hours)

⏰ Scheduler

→

📨 NATS

→

⚙️ Worker

→

🌐 APIs

→

🗄️ DB

Flow 2: User Requests Analysis

👤 User

→

🚪 API

→

🎖️ Commander

→

📊+✍️+💰 Parallel

→

📋 Result

Flow 3: Authentication

📝 Register

→

🔒 bcrypt 12

→

🎫 JWT 15m

→

🔄 Refresh 7d

Flow 4: AUTOPILOT End-to-End

🎨 Generate

→

🚀 Deploy

→

📚 Learn

→

⚡ Optimize

→

📊 Report

See security and compliance (5 layers)

Layer 1 Edge: Cloudflare DDoS + WAF + Bot Filter + Geo Block

Layer 2 Transport: TLS 1.2+ Full Strict + HSTS + Let's Encrypt

Layer 3 Application: JWT HS256 (15min) + bcrypt 12 + Rate Limiting + RBAC (Owner/Admin/Member/Viewer)

Layer 4 Data: AES-256-GCM (OAuth tokens) + Input Sanitization + GDPR/PDPA Compliance

Layer 5 AI Safety: Confidence thresholds (≥90% OK, 70-90% Warn, <70% Block) + Budget limits + Human-in-the-Loop

Hot Air Balloons at Sunset — visual break before development roadmap

Almost there — the development roadmap and how to get involved are next.

What does the 6-phase, 30-day development plan look like?

The entire project is split into 6 phases. Phase 1 is 95% done — the rest is waiting for the team to jump in.

✅ Phase 1: Foundation

Days 1-5 | Status: 95% complete

Monorepo + Docker + Next.js + Go API + DB Schema + SQLC + Auth (JWT + bcrypt)

Phase 2: Core Backend

Days 6-12 | Status: Pending

Platform Connectors (Google, Meta, TikTok) + Data Ingestion + Campaign Service + Analytics

Phase 3: AI Agents

Days 13-20 | Status: Pending

LLM Router + Agent Framework + 9 Core Agents — the AI can analyze and generate content

Phase 4: Autopilot

Days 21-25 | Status: Pending

Creative Generator (Picasso) + Thompson Sampling (Darwin) + Autopilot Workflow

Phase 5: Frontend

Days 10-25 (parallel with backend) | Status: Pending

Design System + Dashboard + Campaign UI + Autopilot UI + Settings

Phase 6: Polish

Days 26-30 | Status: Pending

Testing (Unit + E2E) + Security Audit + Monitoring (Prometheus + Grafana) + Deploy

Cursor Prompt: "Set up a Temporal workflow that orchestrates 3 AI agents in parallel — analyst, creative writer, and budget optimizer — then merges their results"

What kind of help does the project need?

If you're interested, pick the task that matches your skills — there are 3 priority levels to choose from.

Priority	Task	Skills Needed	Phase
🔴 High	Platform Connectors (Google, Meta, TikTok)	Go + OAuth2 + REST APIs	2
🔴 High	Data Ingestion Pipeline	Go + NATS + TimescaleDB	2
🔴 High	LLM Router + Agent Framework	Go + AI APIs + Temporal	3
🟡 Medium	Dashboard UI	Next.js + Recharts + TanStack Query	5
🟡 Medium	Campaign Management UI	Next.js + Shadcn/UI	5
🟡 Medium	AUTOPILOT UI + Wizard	Next.js + Complex Forms	5
🟢 Normal	Thompson Sampling Engine	Go + Statistics + MAB	4
🟢 Normal	Testing (Unit + E2E)	Go Test + Playwright + k6	6
🟢 Normal	Monitoring Setup	Prometheus + Grafana + Loki	6

How do you actually get started building?

Type these prompts in Cursor AI — the AI sets everything up for you. No manual terminal work needed.

01Clone repo + Install

Cursor Prompt: "Clone the AdsPilot repo and install all dependencies for me"

02Start Infrastructure (Docker)

Cursor Prompt: "Start all Docker services for AdsPilot — postgres, redis, nats, temporal"

03Setup Environment + Run Migrations

Cursor Prompt: "Copy .env.example to .env and run all database migrations"

04Start API + Frontend

Cursor Prompt: "Start the Go API server and the Next.js frontend"

API: http://localhost:8080/health | Frontend: http://localhost:3000

See the actual commands (if you prefer typing them yourself)

git clone <repo-url> && cd adspilot && pnpm install

    docker-compose -f infra/docker/docker-compose.yaml up -d

    cp .env.example .env

    docker exec -i adspilot-postgres psql -U adspilot -d adspilot < packages/db/migrations/001_init.up.sql

    cd apps/api && go run ./cmd/server/

    cd apps/web && pnpm dev

See the full project structure

adspilot/

    ├── apps/

    │   ├── web/             # Next.js 15 Frontend

    │   └── api/             # Go Backend API

    ├── services/

    │   ├── agents/          # AI Agent Service (9 agents + LLM Router)

    │   ├── autopilot/       # AUTOPILOT Engine (Picasso + Darwin + Sherlock)

    │   ├── connectors/      # Platform Connectors (Google, Meta, TikTok...)

    │   └── worker/          # Background Workers (Sync, Scheduler)

    ├── packages/

    │   ├── db/              # Migrations + SQLC queries + Schema

    │   └── shared/          # Shared utilities

    └── infra/               # Docker + Kubernetes

See the full database schema (11 tables)

PostgreSQL 16 (Core): organizations, users, ad_accounts, campaigns, ad_sets, ads, autopilot_sessions, autopilot_variations, autopilot_decisions, agent_tasks, reports

TimescaleDB (Metrics): ad_metrics (hypertable, 1-day chunks, compress after 7 days, retain 2 years) + continuous aggregates: hourly + daily

Key Fields: Every table has UUID PK + org_id FK + created_at/updated_at | autopilot_variations has alpha/beta for Thompson Sampling + ctr/cvr/roas/score

See all API endpoints

Auth: POST /auth/register, /auth/login, /auth/refresh, /auth/logout

Campaigns: GET /campaigns, GET /campaigns/:id, PATCH /campaigns/:id/status, POST /campaigns/bulk-status, GET /campaigns/:id/ads

Analytics: GET /analytics/dashboard, /analytics/timeseries, /analytics/top-campaigns, /analytics/platforms

AUTOPILOT: GET+POST /autopilot/sessions, GET /sessions/:id, POST /sessions/:id/launch, GET /sessions/:id/variations, GET /sessions/:id/decisions

AI Agents: POST /agents/chat, GET /agents/tasks, GET /agents/tasks/:id

Settings: GET+PATCH /settings/profile, GET /settings/connections, POST /settings/connections/:platform/oauth, GET /settings/organization

Dubai Skyline Reflection — premium visual break

💰

Last section — Pricing

Starting at $28/month vs $550+ for agencies — affordable for any business size.

What do the pricing plans look like?

Starter

$28

Ad Spend ≤ $1,400
1 Ad Account
Analyst only
AUTOPILOT ❌

Growth

$84

Ad Spend ≤ $5,600
3 Ad Accounts
3 Agents
AUTOPILOT Manual

Pro ⭐

$280

Ad Spend ≤ $28,000
10 Ad Accounts
All 9 Agents
AUTOPILOT Full Auto

Enterprise

Custom

Unlimited
Unlimited
All Agents
Custom AUTOPILOT

AI is about to change how every business runs ads — the only question is whether you'll build the tool or pay someone else to use one.

Want to help build this?

Check the Issues on GitHub — there are good-first-issue labels for newcomers
Read the Docs in Docs/ — every phase has detailed specifications
Send a PR — every contribution is welcome