Skip to main content
Start a Project
Tafakor Pardazan AI Assistant filling in a purchase invoice form inside the accounting app, shown with its Persian and English assistant panels
Internal products

Tafakor Pardazan AI Assistant: A Digital Employee for Business Software

An AI-powered digital employee that turns a natural-language request into real work in the software, using a visible mouse pointer on the interface of Tafakor Pardazan Accounting.

  • Docker
  • Go
  • LLM
  • SQLite
  • TypeScript

Tafakor Pardazan’s AI assistant works as a digital employee: it turns a Persian request into real work inside accounting software.

AI assistant: details and key features

The problem: accounting work is full of repetitive steps

Every accounting entry starts with a simple sentence, like “register this purchase invoice”, but finishing it means opening several screens, finding or creating the supplier and filling in dozens of fields. Chatbots can explain these steps, but they do not carry them out.

The solution: a digital employee inside the software

Tafakor Pardazan AI Assistant is an independent product that understands a natural request, turns it into a reviewable plan, and does the work in the real interface of the software. Read the product page

It understands

It reads Persian text, documents (receipts, invoices, bank slips) and, where enabled, voice, identifies the business domain, and matches the request to the company’s real parties and accounts.

It works on screen

It moves a visible mouse pointer across the real interface, clicks and fills in fields; or, where supported and authorized, it executes directly through the application’s services.

Controlled and reviewable

It asks for approval before registering (the default), reads the result back from the application, and records every action with before and after values.

The mouse pointer: what sets an agent apart from a chatbot

In this project the mouse is visible: the user can follow the pointer and the assistant’s actions on screen. This is the default, and direct execution is the second option where the application supports it. How a task runs depends on the workflow, the user’s permissions and the available tools.

Real-World Product Demonstration
A real screen recording: registering a purchase invoice in Tafakor Pardazan Accounting. The operating system’s file-selection window is blurred for privacy. View it in the Videos section →

Inside Tafakor Pardazan Accounting

The assistant’s current deployment is in Tafakor Pardazan Accounting, a web-based double-entry accounting system. The assistant and the accounting system are separate products: the assistant is an independent product that now works inside this software, and its architecture is designed to extend to other applications (no other integration is offered today).

The assistant's proposed plan for registering a purchase invoice, with the options
The assistant’s proposed plan for registering a purchase invoice, ready for approval.
The
The assistant fills in the “New person” form on screen.
The new purchase invoice form with the supplier selected; the assistant's pointer is on the notes field as it enters the text
The purchase invoice form with the assistant’s pointer.
The line item and amount of the purchase invoice entered by the assistant, with its pointer on the tax field
The line item and amount entered on screen.
The Tafakor AI panel in English, with the
The assistant panel, English interface
The Tafakor AI panel in Tafakor Pardazan Accounting in Persian, with starter suggestions and the
The assistant panel, Persian interface

Technical details (from the project’s code)

  • Architecture: a browser widget (TypeScript, Shadow DOM) and a Go service with its own SQLite store; the service calls the accounting API with the user’s own session and permissions.
  • Models: calls are made server-side through OpenRouter, and the key never reaches the browser.
  • UI agent: language-independent tags and an app map; targeting by tags, then accessibility, then vision; drivers for React inputs, combo boxes and the Jalali date field; every step is verified.
  • Safety: a deterministic policy engine (not the model) decides the approval mode; approval tokens are single-use and bound to the exact request; tools are an allowlist; Undo and an activity log.
  • Coverage: more than 200 application operations identified from the code; sensitive operations always require human approval.
  • Testing: service unit and integration tests, widget tests and browser end-to-end tests.
Product page Tafakor Pardazan Accounting Video in the Videos section

Related projects and services

Related work: TP Call and TP Accounting. This work belongs to our Artificial Intelligence services. To discuss a similar project, contact Tafakor Pardazan.

Order Request

Order This Project

Tell us what you need. Our team will review it and contact you with a proposal.

    Have a project in mind?

    From idea to prototype and product, we're with you.

    Start a Project View All Projects