AI Chatbot Platform
A control plane that deploys a retrieval-augmented chatbot into each customer's own AWS account or Azure subscription, one isolated stack per tenant. The platform never sees chat traffic, documents or logs, and holds no credential for any customer cloud: the customer's one-click setup creates an identity that trusts only that chatbot's GitHub Actions runs, and every deploy signs in with a token GitHub issues for the run.
It is a Next.js application backed by Postgres. Deployments run as GitHub Actions workflows that apply Terraform in the customer's cloud, with the Terraform state kept there too.
Documentation
| Document | Covers |
|---|
ARCHITECTURE.md | The control-plane / data-plane boundary, per-tenant infrastructure on both clouds, secret and credential flows, data model |
DOCS.md | Reference: routes, database schema, onboarding flow, workflows, Terraform variables, environment variables, cost model |
docs/DEPLOYMENT-AWS-VS-AZURE.md | A deployment step by step on each cloud, and how the two paths compare |
CLIENT-DEPLOYMENT-GUIDE.md | What a customer prepares in their own cloud, and what to expect afterwards |
DOCUMENT-MANAGEMENT.md | Uploading and deleting knowledge-base documents without the platform holding a storage credential |
CHATBOT-LOGIC.md | What the deployed chatbot does: indexing, retrieval, prompting, configuration |
SECURITY.md | Credential inventory, the trust model on each cloud, and known security limitations |
LIMITATIONS.md | Known architectural limitations, and why each exists |
EVALUATION.md | How the platform's claims are to be evaluated |
docs/figures/ | The figures the documents refer to |
The chatbot backend itself lives in the separate ai-chatbot/ai-backend repository.
Running locally
npm install- Copy
.env.example to .env.local and fill it in. Each variable is explained there; the AWS ones assume the one-time setup in infra/platform/.
- Apply the database schema to your Postgres database (
npm run db:migrate).
npm run dev, then open http://localhost:3000.
The deploy workflows reach the platform at the PLATFORM_BASE_URL repository secret, which must be publicly reachable. Browser uploads to a tenant's storage are accepted only from that address and the optional EXTRA_CORS_ORIGIN repository variable, so a platform run at localhost:3000 needs EXTRA_CORS_ORIGIN=http://localhost:3000, followed by a redeploy of each tenant (LIMITATIONS.md #7).
Tests
npm test runs the Vitest suite. Much of it checks the infrastructure code as source — the workflows, Terraform and setup templates — for the security properties the documents claim.
The security evaluations run locally, create no cloud resources and cost nothing. Their method and results are described in EVALUATION.md §5:
| Command | What it runs |
|---|
npm run evaluate:security:e2e | Four end-to-end attack scenarios (SEC-E2E-01 to 04), each paired with controls |
npm run evaluate:security:mutants | Mutation testing: removes each security control in turn and checks a scenario catches it |
npm run evaluate:security:concurrency | The races repeated against PostgreSQL in Docker (needs Docker) |
