I build configurable software for complex enterprise workflows.

Local-first desktop expense tracker — no account, no server, no sync.

• Models every money movement as one row in a single table — spend, income, self-transfer and card charge differ only by direction and which source column is set; amounts are positive integers in minor units.
• Balances are derived in SQL from an opening balance plus every transaction that touched the account, so no running total can drift out of sync with the ledger.
• Five screens over the same period picker: analytics with period-over-period deltas and CSS-box charts (no charting library), plus a rules-based report generator that works with no model configured.
• Optional Ollama integration categorises transactions in batches against a closed 14-category list and rewrites the report prose; the model writes sentences, never figures — every number is recomputed from the ledger.
• Ships via GitHub Actions for macOS/Linux/Windows with minisign-signed auto-update; disk and network access live in Rust commands, the webview never fetches.

Offline gamified home-workout desktop app, plus the reactive store it is built on.

• Deterministic plan generation: the same profile, catalog and locations always produce an identical weekly plan, and the generator version stamped on each snapshot forces an explicit regeneration instead of a silent replay.
• Personalization keyed to a training ground — each place carries its own equipment set and exclusions, so eligibility is computed per location rather than from global checkboxes.
• XP, level, rank, streak and weekly count are derived from the ordered completion history; an INSERT OR IGNORE on the date primary key makes a repeat claim idempotent, with no mutable summary row to drift.
• Every persisted value passes a runtime shape guard that fails closed — a single corrupt row is skipped instead of blocking boot.
• Published flex-state as a framework-agnostic store (Object.is dedupe, detachable methods) with a React 19 binding over useSyncExternalStore; tested with Vitest and Node’s native test runner.

Local-first study-plan tracker for macOS, Android and iOS — no account, no server.

• Two clients over one model: @growth-os/learning is consumed as TypeScript source with no build step, so the Tauri 2 desktop app and the Expo React Native app share the same types, validators and eight bundled courses.
• Desktop progress is dated completions in SQLite — one row per finished task, so streaks and totals are derived from the ordered history and re-ticking never rewrites the original date.
• Task chat and day summaries call Ollama from Rust, never the webview, so the API key never reaches the frontend; the prompt is rebuilt from the current path on every send.
• Day summaries are written from that day’s own ticks and task conversations, never stored, and appended into a markdown notes folder under a dated heading so no earlier note is overwritten.
• Versioned backup and restore merges into the device and keeps the dates tasks were finished; a file from a newer format version is rejected whole rather than silently losing ticks.