How eBuzz.ai is built
A local-first, agent-native marketplace on AWS. It has chat-first clients, a real-time Deal Room, an event-driven core, and a governed AI agent platform. Security and compliance are designed in from the start: personal data stays encrypted on users' devices, and people approve anything that moves money, publishes content or signs contracts.
Architecture principles
Six decisions shape everything else.
Chat-first, component-rich
The chat stream carries typed UI components (cards, tables, forms). Dashboards hold the full lists.
Local-first data
SQLite on every device syncs with Postgres. PII lives in an end-to-end encrypted vault.
Event-driven core
Orders, offers, games and agents publish events. Durable workflows handle money.
Governed agents
Agents reach services only through MCP tools with least privilege, budgets, evals and human gates.
Protocol-ready
UCP/ACP endpoints let external AI agents shop eBuzz live offers.
Buy, don't build (non-core)
We build Deal Room, Concierge, games and the ledger. Payments, tax, shipping and identity are bought.
Logical architecture
Layers from the device to external services. Security & compliance and observability run through every layer.
Each service owns its Postgres schema and publishes events. There is no cross-service database access. The Ledger (wallet credits, commissions, payouts) is append-only, with double-entry bookkeeping.
The orchestrator (Temporal) runs agent workflows. The runtime calls Claude through the model gateway. Tools are MCP servers backed by domain APIs. Side effects above set limits go to the approval queue.
The sync service replicates allowed rows to each device. Vault columns hold only ciphertext. Fulfilment tokens release a shipping address to one vendor for one order.
Key flow: "Ask for a better deal"
The shopper keeps chatting while vendors respond over several minutes.
Deployment & environments
Everything is deployed through GitOps. Every environment is created from the same Terraform modules.
Tools needed & alternatives
Each category shows the recommended tool, why we chose it, and the alternatives with their pros and cons. Click a row to expand it.
Security architecture
Security is a product feature for eBuzz. Shoppers trust us with purchases, vendors trust us with pricing, and regulators watch AI and gamification closely.
Controls by domain
Secure delivery pipeline
AI & agent security (OWASP Top 10 for LLM Applications)
How each risk is handled for both the operational agents and the build agents.
Threat model (STRIDE, top risks)
Compliance program
What applies to us, why, and when we'll be ready. Counsel should review this before launch.
Fractional CISO from month 0, and a Security Lead hired by month 4. A Risk & Compliance committee meets monthly. An AI governance board approves agent autonomy changes.
Scans run on every PR, DAST weekly, a third-party pentest before launch and yearly, AI red-teaming quarterly, and a public bug bounty 3 months after launch.
About $420k: people ($260k), tooling ($90k), pentest and audits ($70k). That's roughly 8% of the seed round, sized for a trust-sensitive marketplace.