SmallBizPal

An AI-powered team that automates marketing, customer service, and analytics to fuel small business growth.

Comment

SmallBizPal: The Secure AI Workforce for Small Businesses

Inspiration

While testing AI tools at our friend's bakery, we discovered a harsh truth: small businesses are caught between two bad options. They could either:

  • Risk exposing sensitive data using public chatbots ("Our secret recipe leaked in a customer service chat!"), or
  • Pay big fees (~$10k+) for custom enterprise solutions

The breaking point came when another flower shop friend showed us her "AI workflow" – a dozen open ChatGPT tabs and a notebook of copied prompts. We realized: Small businesses don’t need more AI tools – they need an AI team they can trust.

What It Does

SmallBizPal deploys a secure, autonomous AI team that:

  • Guards business secrets (internal strategy vs. public-facing agent)
  • Works 24/7 on the business’s own website via one-line plugin
  • Learns continuously – upload a new menu PDF and it instantly knows the changes
  • Speaks in your brand voice (learns from past customer chats)

Unlike using a ChatGPT like generic tools which is potentially: single leak-prone bucket, only works on OpenAI’s site, forgets after chat ends, we plan to improve experience across these areas. SmallBizPal provides Data Security with separation, deployment is embedded on your website and remember business context

How We Built It

Core Architecture: We built SmallBizPal on a sophisticated but secure multi-agent foundation using the Google Agent Development Kit (ADK). Our design is centered on a Dual-Application Model to enforce a strict security boundary:

  • A private, admin-facing app for managing business secrets.
  • A lightweight, public-facing app for safe customer interaction.

For production, the apps communicate via Google's secure A2A (Agent-to-Agent) protocol, providing a zero-trust architecture.

  • Frontend: Next.js + Tailwind (admin), Web Components (widget)
  • Backend: FastAPI (OAuth2 scopes; strict internal/public separation) → Cloud Run
  • AI Stack: Gemini models + Python ADK (agent orchestration)

Challenges we ran into

  1. The Leaky Bucket Problem Issue: Early versions sometimes leaked pricing strategies Solution: Implemented dual-vector RAG with semantic similarity blocking

  2. The "I Hate Chatbots" Bias Issue: 68% of small biz owners rejected initial demos ("Chatbots feel fake") Breakthrough: Added human handoff triggers ("Press 0 to talk to Sarah")

  3. PDF Chaos Nightmare: One bakery uploaded a 40-page PDF with recipes scattered everywhere Innovation: Built a layout-aware parser that understands menus/price lists

Accomplishments that we're proud of

We started building our AI explorations since more than a year and incrementally build lego pieces for bigger vision. First one started back in Oct 2024 with a hackathon winning project TagFi around 'verified user data for enhancing model quality'. Then we presented our AgentConnect framework on readytensor.ai and won first price as 'Best overall project' among participants from across 70+ countries. Third lego piece and our win was our agent to agent payment project using Coinbase AgentKit. These incrementally helped us to now combine our lego pieces and create SmallBizPal solution.

This solution also have accomplishments that we feel proud of:

  1. Building a Truly Secure AI System: We successfully designed and implemented a dual-application architecture with a KB-Proxy agent that acts as a secure gateway. This solves the "leaky bucket" problem from first principles, ensuring a small business's private data is never exposed to the public-facing agent.
  2. Creating an Orchestrated, Autonomous Workforce: We went beyond a single chatbot and created a true multi-agent team. Our Orchestrator agent intelligently coordinates tasks among specialized agents for discovery, marketing, and reporting, which directly addresses the "too many tabs" workflow chaos we saw business owners facing.
  3. Designing for Production from Day One: We built SmallBizPal with a clear and practical path to a real-world deployment. The architecture is ready for Google Cloud Run and the agent communication is designed to seamlessly transition from local development to a secure, distributed A2A protocol, making it more than just a hackathon prototype.

What we learned

  • AI personalisation is ready to be sellable
  • Huge percentage of knowledge comes from unstructured docs and chats
  • Owners care about tone more than tech

What's next

"Emergency button": Detect/defuse angry customers before they leave bad reviews Localisation pilot: Spanish + Tagalog support for immigrant-owned businesses Storage: Integrate storage for docs, images, files and chats for future intelligence and audits

Built With

Share this project:

Updates