Skip to content
← All work
SYS_PRIVATEAIACTIVEAI_Agent

Private Business Agent

On-premise AI agent for Bharat PolyWorld, answers business questions on Telegram, WhatsApp and the company app, running Qwen on a Mac in the client's own office.

Private Business Agent: AI_Agent project by SoulPlex Studios

Overview

Built alongside the Bharat PolyWorld operations system. The owner asks a plain question, which purchase orders are pending, how much granule a unit made this month, which supplier sent the most low-grade scrap, who still owes money, on Telegram, WhatsApp or inside the web and mobile app, and gets an answer built from the company's own records, with the sources listed underneath. The model (Qwen 3.5 4B through Ollama) and the retrieval layer run on a Mac M2 in the client's office, not on a third-party AI cloud, so no per-message AI bill and no company data sent outside, and reach the business database through a dedicated read-only account that cannot change anything and cannot see bank details, PAN, phone numbers or login tokens. Only the owner can use it, and it can be switched off from the app in one tap.

Key features

  • One agent across Telegram, WhatsApp, web app and mobile app
  • RAG over live operations data, with sources shown for every answer
  • Self-hosted on a Mac M2 with Ollama, no cloud AI bill, data stays in-house
  • Read-only database role: writes refused, 15-second query limit
  • Sensitive fields excluded: bank, PAN, phone numbers, tokens
  • Owner-only access with an in-app on/off switch

Pipeline

  1. 1

    Owner asks a question on Telegram, WhatsApp or in the app

  2. 2

    Request is checked for role and routed to the on-premise agent

  3. 3

    Agent retrieves matching records over a read-only connection

  4. 4

    Local model writes the answer from those records

  5. 5

    Answer returns to the same chat with its sources