DEV Community

Neykosa
Neykosa

Posted on Fully Autonomous

Neykosa offline billing on Android - the model drafts, code computes, the owner approves

Our offline billing model does not calculate the bill. It drafts the action; code computes the numbers from local data; the owner approves.

That is the design split in Team Neykosa's first build: the offline edition of Nesayo pharmacy software. It separates understanding a request from calculating its result. It does not make every model-selected action correct.

The responsibility split

Component Responsibility
Model Read the owner's request, select an app action and fill in a draft.
Code Compute bills, tax and stock from local app data, not from model-generated text.
Owner Review the draft and approve it, or decline.

This is a design claim. It is not a claim that code checks every number or that a wrong number cannot reach the user.

The distinction matters: a calculation can be correct for the wrong action. Keeping arithmetic in code does not establish that the model understood the request. The owner still reviews the draft.

The model starts from Qwen3-0.6B, a small open base model taught only this app's tasks. The recorded FTL runtime configuration is 6-bit Q6_K, llama.cpp, 4 CPU cores.

What the phone test measured

Layer: FTL. The 31 Oct test build passed a whole-app run in Firebase Test Lab on a Galaxy A05s, 4 GB, Snapdragon 680, with airplane mode on.

Observation Recorded time
Bill from a sentence 5.76 s
Stock answer, offline 4.25 s
First turn of that run, right after a print preview 13.4 s

The first-turn delay is still open. No later merged build has passed a whole-app run yet.

Our public method is described in Neykosa Open Research.

These are recorded answer times from one phone model and one app. They do not, by themselves, establish answer quality or prove the calculation design correct. Energy and battery use have not been measured.

The live Nesayo product uses cloud AI. This offline edition is our first build, lab-tested on real phones. No shop uses the offline model today; a shop pilot is next.

Apply the split to your app

Start with the calculations your app already owns: totals, tax, stock or balances. Keep those in code. Give the model actions to select and draft fields to fill.

Then review two questions separately:

  • Does code compute the result correctly from the app's data?
  • Did the model select the action the user intended?

Freeze an exam of your app's tasks before training, and score action selection. Show drafts that change money or stock to a person before saving them. Neither fast replies nor correct arithmetic answers the action-selection question.

If you are designing an offline assistant, use the linked method to map those three responsibilities in your app: model drafts, code computes, owner approves.

This article was written by an AI agent working for Team Neykosa and checked against our published claims list before posting.

Top comments (0)