Offline AI Desktop Product

What it does

A cross-platform desktop application that automates trial-balance account mapping and financial-statement reconciliation for an air-gapped enterprise audit client — no internet, no cloud APIs, no telemetry. Shipped solo, from spec to signed installer, in two weeks. 99.4% mapping accuracy on the validation set with the AI disabled, and reconciliation reproduced the audit team's workbook line for line.

Architecture highlights

• Deterministic-first classification ladder: exact match → curated alias vocabulary → fuzzy/token matching → local LLM → flag-for-review — enforcing the audit rule that a no-match flag beats an incorrect match
• Embedded on-device LLM (~3GB GGUF via llama.cpp) as a strict last-resort rung, selected through a 15-probe accuracy-and-latency bake-off across 5 local models with documented failure modes
• Correction learning that stays local: auditor corrections are remembered and applied to future runs — the system improves without any data leaving the client's environment
• Snap-to-vocabulary and balance-sign guardrails on every AI suggestion; match-method provenance carried on every row into the UI and Excel export
• 81 automated tests including an accuracy oracle that gates releases on the validation set

Shipping the last mile

• CI-built Windows and macOS installers (~50MB) — per-user, no admin rights, VDI-friendly
• sha256-verified in-app model download for air-gapped side-loading
• Ed25519-signed offline licensing with no activation server and no phone-home