WarehouseLens
SaaS / Logistics / AIA deployed warehouse operations platform running illustrative data - a natural-language copilot, demand forecasting, and warehouse-scoped RBAC. Built solo.
View Live SiteWarehouseLens: Screenshots not published yet.
In progress
A deployed, multi-tenant warehouse operations platform covering inventory, procurement, and outbound logistics. The current deployment uses illustrative data rather than active customer operations, with a Next.js BFF over a FastAPI backend, a natural-language copilot, Prophet/XGBoost demand forecasting, and warehouse-scoped permissions.
Background
Warehouse teams often need inventory, supplier performance, and outbound status brought into one operational view. WarehouseLens is a deployed reference implementation that demonstrates this model with illustrative data rather than active customer operations. The rule the whole design rests on is that the planner never generates SQL: it selects one approved tool, that tool runs a fixed parameterized query, and a second model call turns the structured result into prose. The platform includes a two-level identity model, 6 roles across 21 permissions with warehouse-scope enforcement, and a background worker that keeps aggregation and reservation recompute off the request path.
What it has to get right
Letting a language model answer operational questions without writing its own queries or reading across a tenant or warehouse boundary, while keeping a write-heavy API responsive and the forecasts worth acting on.
Solution
Built a LangGraph plan/execute/synthesize graph where the planner only ever picks a tool, never writes a query, and permission checks run inside the agent so an out-of-scope question is refused rather than answered. Paired it with a Next.js BFF that keeps tokens out of client JS, Keycloak JWTs verified against a cached JWKS endpoint, a Prophet and XGBoost forecasting layer validated by a rolling-fold backtest harness, and a background worker for asynchronous aggregation.
Tech Stack
Related Service
WarehouseLens enforces tenant and warehouse scope everywhere - including inside the copilot, which refuses a question about a warehouse the caller is not assigned to rather than answering it. Designing permissions that hold at every layer is the core of this engagement.
See Multi-Tenant SaaS & Platform EngineeringProject Profile
Personal project, illustrative data
Evidence
- Access control
- 6 roles, 21 permissions, warehouse-scoped
- Copilot
- Tool routing; the planner writes no SQL
- Forecasting
- Prophet and XGBoost, rolling-fold backtest
- Test suite
- 155 tests and an evaluation harnessTested
- Deployment
- Illustrative data, not customer operationsDemo
- Tested
- Evidenced by the test or backtest suite, not by production traffic.
- Demo
- Deployed and reachable, but running illustrative rather than customer data.
- No label
- An implementation fact: this exists in the system.
Links
Interested in building something similar?
Get in touch