Stock Analyzer

Local models for private context

Sentiment classification and general reasoning run against locally hosted models rather than a hosted API.

The reason is not cost — it is that a prompt containing a live portfolio is a document you do not want to hand to a third party. Where a frontier model genuinely helps, the interface assembles the context, places it on the clipboard and opens an existing browser session, so the capability is available without the application ever holding a key.

A backend split by concern

Data acquisition, quantitative analytics, strategy algorithms, backtesting, portfolio state and the API surface are separate modules with separate dependency sets — the machine-learning requirements install independently, so a contributor working on the API never compiles a deep-learning stack to run the tests.

What it proved, and what it did not

It proved the product: the pane layout, the local-AI posture, the broker loop and the research workflow all held up in daily use.

It also proved the ceiling. A web runtime wrapped for desktop cannot carry a live chart surface, an embedded terminal and a 3D view simultaneously without the frame budget collapsing — which is the finding that justified a full native rewrite rather than another round of optimisation. Publishing that plainly is more useful than presenting the successor as a natural evolution.

Secrets are an architectural concern

Broker credentials and service keys are synchronised out of the working tree into a vault rather than living in files beside the code, and the repository carries an explicit audit step before anything is committed.

Treating credential handling as part of the architecture — rather than as a gitignore entry and good intentions — is what makes a single-operator trading system safe to keep in version control at all.