Can a local LLM build a real app? (FinTrack)
Notes from FinTrack — handing the first build to a Gemma model running on local hardware, then hardening it with Claude.
FinTrack started as a question, not a product: how far can a model running entirely on local hardware get on a real full-stack app?
The setup
I ran a Gemma 4 26B a4b QAT model locally through LM Studio's locally hosted server (my own macbook pro m2 max) (127.0.0.1:1234) and pointed OpenCode at it in build mode. It generated the Next.js app piece by piece — a theme toggle, a transaction form, smart income/expense categories, dark mode — narrating each change as it went.
Where it broke
It got surprisingly far, and then it didn't. Mid-build the site collapsed into what I can only call recovery mode: bare HTML, no styling, no charts — "everything in basic mode now, no visualization." Watching a local model confidently rewrite files while the app quietly fell apart was the most instructive part of the whole experiment.
The fix
I handed the broken codebase to Claude to review and harden. It restored the UI, fixed the multi-currency logic (IDR / USD / JPY with daily FX conversion), and brought back the trend charts and category donut. The experiment became something that actually runs.
What I took away
- Local models can scaffold a real app shockingly fast.
- The gap shows up in coherence — holding a growing codebase together, catching edge cases, not breaking what already worked.
- Pairing a local model (cheap, private, fast first drafts) with a frontier model (review and hardening) is a genuinely useful workflow, not a gimmick.