SMB Growth Agent

Command Centre

Inspiration

Salon owners have customer data, but often spot churn, missed rebooking, and festival opportunities too late. We built Salon Growth Agent to act like a proactive Growth Officer: identify what changed, recommend the next best action, and keep the owner in control.

What it does

The agent reads salon CRM data, identifies customer cohorts such as service-cycle due and at risk, and proposes up to three campaigns: rebooking, win-back, and seasonal opportunities like Navratri.

Gemini investigates privacy-safe business evidence, creates a structured growth plan, and answers owner follow-up questions on Telegram. The dashboard shows the evidence, campaign risks, validation result, and approval queue.

How we built it

Python and SQLite handle CRM ingestion, cohort calculations, opportunity ranking, eligibility, campaign state, and audit history locally.

Gemini 3.5 Flash on Vertex AI chooses from allow-listed tools and works only with sanitized aggregate evidence. A Gemini critic and deterministic validator check every plan before campaigns enter the approval queue.

The product includes a local dashboard, an owner-only Telegram bot, and an authenticated Cloud Run notification relay.

Challenges and learnings

The hardest part was making Gemini investigative without giving it control over customer targeting or campaign delivery. We solved this with strict structured outputs, privacy-safe tools, deterministic validation, and two-stage owner approval.

We learned that an AI agent needs more than a model and tools. It needs durable memory, explicit contracts, visible evidence, well-defined ownership of facts, and clear terminal actions. Gemini is valuable for deciding what to inspect, connecting evidence, forming hypotheses, and explaining trade-offs. Deterministic systems are better for customer eligibility, policy, financial logic, state changes, validation, and approval.

What’s next

Next steps include consent and channel-preference ingestion, messaging integrations, frequency caps, outcome tracking and running A/B marketing campaign experiments.