Experiment · 2026
FinTrack
An experiment: can a locally-run LLM build a full-stack app?
A deliberate experiment: how far can a model running entirely on local hardware get on a real full-stack app? The first build was generated by a local Gemma model — running through LM Studio's local server and driven by OpenCode in build mode — then reviewed and hardened by Claude. The result is a working multi-currency expense tracker, and an honest look at where local models shine and where they stall.
The question
Can a model running entirely on local hardware ship something real — not a toy — end to end?
The build
Gemma ran locally via LM Studio (server at 127.0.0.1:1234), with OpenCode generating the Next.js app component by component: theme toggle, transaction form, smart categories, dark mode.
Where it broke
Mid-build the app collapsed into a bare-HTML “recovery mode” — no styling, no charts. A vivid reminder that local models can scaffold fast but struggle to hold a real app together.
The fix
Claude reviewed and hardened the codebase — restoring the UI, multi-currency logic, and charts — turning the experiment into something that actually runs.
Stack