AdPilot

Instant Google Ads audit + WhatsApp monitoring for small businesses.

You don't come to us. We come to you — on WhatsApp, in plain language, only when it matters.

Most small businesses running Google Ads have no time to log into a dashboard and no interest in learning what "CPA" means. So money leaks — on keywords that never convert, on budgets left unbalanced, on nobody watching the account between setup and the bill. Existing tools (Optmyzr, Adalysis, WordStream) assume you already track conversions, already log in daily, and can pay $100+/month. AdPilot assumes none of that.

Live demo: https://ad-pilot-gamma.vercel.app · API: https://adpilot-api-lv8o.onrender.com/health (free-tier backend — the first request after a while takes ~50s to wake)

---

How it works

  Google Ads CSV export
          │
          ▼
   Upload  ──►  Rule engine  ──►  LLM explainer  ──►  Instant audit
 (no login)     (waste, scale,     (plain language,     "You've wasted ₹2,200
                 negatives)         number-checked)       this month" + ranked
                                                          actions, EN / हिंदी)
          │
          ▼  (opt in — only after you've seen the value)
   WhatsApp  ◄──►  you
   • alert only when something meaningfully changes
   • reply PAUSE / SCALE / DETAILS to act (execution simulated for the demo)

The demo, start to finish

1. Upload sample_data/sample_google_ads_report.csv — no signup.
2. See the headline: ₹2,200 wasted this month, and 8 ranked recommendations ("Turn off cheap shoes online — ₹1,240", …). Toggle हिंदी.
3. Open the budget plan: cost per sale ₹170 → ₹90, and a suggested split across campaigns.
4. Enter a WhatsApp number → opt in → a confirmation message arrives.
5. When a new problem appears, AdPilot messages you. Reply PAUSE → it confirms the (simulated) action and the estimated saving.

---

What's in the box (MVP scope)

#	Feature	Where
1	Instant audit — CSV in, headline waste number + ranked actions out, no login	backend/app/engine/rules.py, frontend/app/audit
2	Rule-based recommendation engine (PAUSE / SCALE / NEGATIVE / low-CTR) — no ML	backend/app/engine/rules.py
3	LLM explainer — structured JSON in, plain sentences out; every number is verified against the input or a template is used	backend/app/engine/explainer.py
4	WhatsApp delivery (Twilio) — sent only on a meaningful change, never on a timer	backend/app/scheduler/jobs.py
5	Reply-to-act loop — PAUSE / DETAILS / SCALE; the Google Ads mutation is simulated and labelled as such	backend/app/services/reply_service.py
6	Chrome extension — thin MV3 popup over the same API (no page scraping, no OAuth)	extension/
+	Budget optimizer + simulator	backend/app/engine/{budget,simulation}.py, frontend/app/plan

Bilingual: English + Hindi, plumbed through upload → analyze → explainer → WhatsApp. No PPC jargon anywhere a business owner reads — enforced by a test.

Explicitly out of scope (48-hour build): live Google Ads OAuth / API writes, any ML model, other ad platforms, billing, user accounts with passwords.

---

Tech

Layer	Choice
Frontend	Next.js (App Router) · React · Tailwind
Backend	Python · FastAPI · pandas · SQLAlchemy
Database	PostgreSQL (Supabase)
LLM	Anthropic Claude (claude-sonnet-5) — explanation layer only, never the maths
Messaging	Twilio WhatsApp Sandbox (outbound + inbound webhook)
Extension	Plain HTML/JS, Manifest V3
Deploy	Frontend → Vercel · Backend → Render · DB → Supabase

Numbers are computed by rules in Python and checked; the LLM only rewrites them into sentences. No figure a user sees is ever hallucinated.

---

Repository layout

backend/       FastAPI service — engine/ (rules, metrics, explainer, budget,
               simulation), routes/, services/, scheduler/, models/, tests/
frontend/      Next.js app — landing, upload, /audit/[id], /plan/[id]
extension/     Manifest V3 popup
sample_data/   sample_google_ads_report.csv

Design docs: 01_PROJECT_REQUIREMENTS · 02_ARCHITECTURE · 03_RULES (binding behaviour) · 04_PHASES · 05_DESIGN

---

Run it locally

# backend  →  http://127.0.0.1:8000  (docs at /docs)
cd backend
python -m venv .venv && .venv/Scripts/python -m pip install -r requirements.txt
cp .env.example .env
.venv/Scripts/python -m uvicorn app.main:app --reload
.venv/Scripts/python -m pytest        # 89 tests

# frontend  →  http://localhost:3000
cd frontend
npm install
cp .env.local.example .env.local      # NEXT_PUBLIC_API_BASE_URL
npm run dev

With no .env, the backend uses a local SQLite file and the audit works with no external services (WhatsApp / LLM degrade to templates). Details in backend/README.md, frontend/README.md, extension/README.md.

Deploying: DEPLOY.md. External-service setup: SETUP.md.

---

Status

All build phases complete (0–9). The reply-to-act execution is simulated for the demo — everything else runs on real infrastructure.

Demo-day note: the Twilio sandbox only sends free-form messages within 24h of the recipient's last message to it. Message the sandbox number from the demo phone right before presenting.