Case study · · 4 min read

Building Orbyte: an AI operations manager that asks before it acts

How ASIIFDEV built Orbyte, an AI operating system for software agencies that connects business records, prepares operational actions, and keeps high-impact decisions under human approval.

Muhammad Syaiful Anwar, Founder & Lead Engineer PT Asiifdev Digital Agency

Muhammad Syaiful Anwar

Founder & Lead Engineer, ASIIFDEV

Orbyte is an AI business operating system developed by ASIIFDEV for software agencies and startups. It brings clients, products, subscriptions, invoices, leads, tasks, documents, meetings, deployments and marketing work into one workspace. Instead of acting as another chat window, its assistant works with typed tools against the organisation's records: it reads, researches, drafts and prepares actions for a person to review.

Orbyte is an ASIIFDEV-developed product, not a client case study. The product was built for the Startup track of the Meta Global AI Developer Hackathon and uses Meta Muse models for assistant, multimodal and creative workflows.

The operational problem: work is scattered, follow-through is manual

Small software agencies often manage client relationships, recurring subscriptions, invoices, proposals, project tasks and deployments across separate tools. The hard part is not simply storing the records. It is noticing what needs attention, understanding the context, and taking the next step without losing control of money, commitments or production systems.

Orbyte was designed around that operational gap: connect the records, surface exceptions, and let AI prepare the next step while keeping consequential decisions reviewable.

An assistant connected to business tools, not just a prompt box

The assistant uses typed tools across the product's modules. Read tools retrieve business facts, draft tools prepare items such as proposals, invoices, tasks and marketing copy, and action tools request changes through the approval flow. The repository documents more than 80 tools across clients, products, packages, deployments, billing, pipeline, tasks, documents, meetings, marketing, connections and AI usage.

The assistant can also research the web, read company websites and process uploaded documents, images, audio and video. A scheduler works in the background to find operational events such as overdue invoices, upcoming renewals, late tasks and deployment issues, then creates findings with an explanation and a suggested next step.

The key design decision: the model proposes, the system enforces

Orbyte separates language generation from business-critical calculations and state changes. The model can explain a balance or draft a proposal, but ordinary application services own invoice totals, taxes, due dates, payment state and ledger balances. AI-generated numbers that do not match the underlying records are rejected.

  • Read: retrieve information from the organisation's records.
  • Draft: prepare a proposed message, document or task without sending or finalising it.
  • Act: request a state change through an approval card that shows the exact arguments.
  • High-impact actions involving money, locked documents or servers require a person to approve them.
  • Every consequential change is recorded in an append-only audit trail; approval permissions and data are checked again when the decision is made.

A practical architecture for trust and control

  1. Business data stays in the application layer. Tenant isolation is enforced in code, and each organisation's records are queried within its own scope.
  2. The model works through typed tools. It does not receive unrestricted database access or free-form shell execution.
  3. Approvals are explicit. The Action Center displays the proposed operation and its exact arguments before a person approves or rejects it.
  4. The service layer performs the change. Calculations, permissions, validation and state transitions remain deterministic application logic.
  5. The audit trail records the result. Changes are attributable and reviewable, while idempotency guards prevent scheduled events from creating duplicate records.

More than chat: billing, meetings, marketing and deployment workflows

  • Billing and subscriptions: recurring renewals can produce invoice drafts; invoices have numbered sequences and locked final documents; payments are represented through an append-only ledger.
  • Meetings: uploaded recordings or Telegram voice notes are transcribed with speaker separation, then turned into structured minutes and proposed follow-up tasks or leads.
  • Marketing Studio: Muse Image creates visuals; captions and scripts are drafted from company data; price and numeric claims are checked against source records before being rendered.
  • Deployment operations: GitHub, SSH and aaPanel connections support visual deployment flows with validation, approvals, execution status and audit history.
  • Telegram: paired users can ask questions, receive findings and review approval cards through the same decision service used by the web app.

Model usage is visible and budgeted

Every model call is logged with its feature, status, token usage, cost and tools used. Daily budgets can be set per feature and calls are blocked when a budget is exhausted. This makes AI usage an operational cost that can be inspected and controlled, rather than an invisible background expense.

What the build demonstrates

The important engineering lesson is that an AI operations product needs more than a capable model. It needs a reliable business data model, strict tenant boundaries, deterministic financial services, explicit permission checks, human approval for risky actions, idempotent background jobs and an audit trail. The model makes the workflow easier to operate; it does not become the source of truth.

Orbyte's repository documents 26 domain modules and more than 80 typed assistant tools. Its README records 1,556 passing automated tests at the 5 November 2026 repository snapshot. These are implementation facts from the project documentation, not claims about customer adoption or business outcomes.

Explore Orbyte, watch the product demo, or read the project repository. If your team is considering an AI workflow that can prepare actions but must keep people in control, talk to ASIIFDEV about the process and risk boundaries first.

Related service

AI automation servicesSee the service →

Want a starting number for your project? Website & app cost estimator.

Working on something similar?

Tell us what you need on WhatsApp or email. We reply with the right questions, not a generic proposal.