INVENTORY / ENGINEERING
A place for everything. An AI that knows its limits.
A personal inventory system that helps you find what you own, remember where it lives, and ask useful questions about it.
Full-stack developer
September 2026 — ongoing
In development
01 / THE PROCESS
The starting point
An inventory is only useful if it reflects how people actually keep things. An item might be in a room, inside a bag, inside another container. Some possessions are individual objects; others, like supplies, have quantities that change.
Inventory brings those physical details into one searchable workspace. The central design question became: how can an assistant help make sense of that information without becoming an unreliable source of truth?
02 / THE PROCESS
Start with the system, then add the AI
I began with a deliberately small architecture: a Next.js interface and API, a user-scoped repository, and Supabase Auth and PostgreSQL. Keeping the app in one deployment made shared validation practical without introducing a separate agent service.
Spaces and Rooms represent locations. Containers are possessions that can hold other possessions. Paths are derived from the hierarchy, so moving a bag does not require rewriting the location of every object inside it.
03 / THE PROCESS
Give the assistant a narrow boundary
Jarvis can search records and reason over a bounded snapshot. It does not receive a database client, SQL tool, or authentication token. Exact requests such as everything in a location are resolved deterministically, rather than relying on a model to claim that a short list is complete.
The first design was read-only. The project later added explicit edit proposals: the model prepares changes, the browser validates them, and the existing authenticated write routes apply them. Packing suggestions never silently change saved inventory.
Structured outputs and verified citation IDs keep answers connected to records that exist. When inference fails, ordinary search remains available with a visible explanation.
04 / THE PROCESS
Make the data harder to break
Ownership is checked at the API and independently enforced with PostgreSQL row-level security. Relationships must belong to the same owner. Hierarchy constraints reject cycles and invalid container relationships.
Versioned writes expose conflicting edits instead of silently overwriting them. Private photos are stored in owner-specific paths and accessed through expiring signed URLs. Vitest, React Testing Library, and PGlite tests cover validation, authentication, database policies, concurrent changes, and AI responses.
05 / THE PROCESS
Where it landed
The public app supports an in-memory sample workspace; signed-in accounts can persist inventory and private photos when the backing services are configured. The repository includes the implementation, migrations, tests, and the decisions behind them.
This is still an ongoing project. There are no user-growth or business-impact claims here. The work so far is about making a useful interaction trustworthy all the way down to storage.
SOURCES & FURTHER READING
Repository and architecture ↗Original design decisions ↗Written from project documentation and my resume. Updated October 2026.
F3 Global
Public-facing pages, reusable interfaces, and protected publishing tools built with a 12-person team.


