An inventory and invoicing workspace for small businesses. I built flexible product schemas, conversational inventory tools, and supplier-document extraction, alongside ARCA electronic invoicing.
These historical architecture notes describe a FastAPI/Astro design and have not been reconciled with the current Next.js codebase. Current product screenshots are available in Selected Work.
Most portfolio projects pick one domain. Stockear spans AI, SaaS infrastructure, and regulatory compliance — simultaneously.
Ask your inventory natural questions.
Behind the scenes, embeddings (VoyageAI + Google GenAI) power semantic search over your product catalog. The agent orchestrates tool calls — CRUD, search, invoice generation — in a single conversational turn.
Per-tenant quotas without leaky abstractions.
The historical design models workspace isolation and subscription tiers, with dependency-layer quota checks before protected business logic.
AFIP-integrated voucher emission.
The historical design describes ARCA/AFIP invoice generation through a hexagonal billing module, with separate adapters for testing and tax-service integration.
Four connected layers — each with clear boundaries and defined responsibilities.
Built with Astro 6 + React islands for maximum performance. Static shell with selective hydration — only interactive components ship JavaScript. Lucia Auth handles session management with secure HTTP-only cookies.
Not every choice is obvious. These architectural decisions explain the why behind Stockear's design.
Argentine tax regulations (AFIP/ARCA) change frequently. The billing module needs to be testable without real API calls and must remain decoupled from the HTTP framework.
Apply ports & adapters (hexagonal) pattern to the billing domain. Domain logic has zero imports from FastAPI, Motor, or arca_arg.
The AI chat interface needs to call multiple backend tools (inventory CRUD, semantic search, invoice generation). Hardcoding tool dispatch would create tight coupling and make adding new capabilities expensive.
Implement a Tool Registry where each tool is a self-describing callable with a schema. The agent reads the registry at runtime and routes LLM-requested tool calls dynamically.
Every API endpoint must respect per-tenant subscription limits. Scattering quota checks in business logic would lead to inconsistencies and bypasses.
Inject a QuotaVerifier as a FastAPI dependency into every protected endpoint. The verifier reads tenant subscription from MongoDB and raises HTTP 429 before business logic executes.
Argentine fiscal compliance requires an immutable log of all transactions, schema changes, and agent actions. The system must be auditable and replayable.
Use MongoDB as an append-only event store. All significant actions (chat messages, invoice emissions, schema changes, quota events) are written as immutable event documents.
From browser to compliance authority — every layer mapped out.
Static Astro shell with selective React hydration. Zero JS for static sections.
Astro deployed to Vercel. SSR pages rendered at edge; static assets globally cached.
Main orchestration layer: agent, tool registry, quota enforcement, business logic.
Each workspace gets an isolated schema. Also acts as event store for audit trail.
Legal invoice emission via hexagonal adapter. Staging and production environments.
Embedding generation for semantic search. Vectors stored in MongoDB for retrieval.
The live application and related source profiles.