Extraction

The platform replaced manual document and image data entry for language-specific teams — a cross-platform desktop app on a self-hosted Supabase/Postgres backend, with encrypted key vaulting, role-based access, and over-the-air updates.

I own it from the first scoping call to production support.

Work routes across five model providers — OpenAI, Google Gemini, Qwen, Vector Engine and local Ollama — behind one interface, so workloads move on price or capability.

Template-driven extraction lets non-technical operators add new document types without engineering involvement, and per-project cost analytics make spend predictable before a run.

Every response is validated against the template's JSON schema before it counts. A failure doesn't just retry — it escalates: a firmer instruction, then a larger output budget, then a different reasoning setting. Long runs checkpoint to disk, so closing the app mid-batch costs nothing.