What is RafayGen? One AI assistant for chat, voice, images and documents

RafayGen — short for "Rafay ka AI", Rafay's AI — is an AI assistant built and operated from Sindh, Pakistan. This article explains what it actually is, how the pieces fit together, and, because it is more useful than another list of features, what it does not do.

The short version: it is a browser-based assistant that combines multi-model chat, a spoken voice mode, image generation, generation of real office files, and Urdu optical character recognition behind a single account. The longer version is worth reading if you are trying to work out whether a general assistant is the right shape of tool for you at all.

The problem it was built around

There is a specific gap that most global AI products leave open, and it is not a technical one. It is that their checkout page assumes an international credit card, their interface assumes you type in English, and their defaults assume a fast connection and a recent phone. Each of those assumptions is reasonable in the markets those products were designed for. Together, in Pakistan, they exclude most of the people who would get the most out of the technology.

So the design constraints were fixed before any of the features were: it has to work in a browser on a mid-range Android phone; it has to accept Roman Urdu as a first-class input rather than as something to be translated away; and it has to be payable without a Visa card. Everything below is downstream of those three.

One assistant, two interfaces

RafayGen presents as two surfaces over one system. The chat workspace is the one you type into: a conversation with markdown, code blocks, tables, generated images and file downloads inline. The voice orb is the one you talk to: press, speak, and hear the answer spoken back while it is still being written.

They are not separate products with a shared logo. They use the same account, the same conversation history and the same underlying model routing, so a topic you were working on by typing is still there when you switch to voice. That is a deliberate choice with a real cost — keeping two interaction models in sync is harder than shipping two apps — and it is made because switching apps to change how you are talking to the same assistant is a strange thing to ask a person to do.

Six agents rather than one model with a size dropdown

Most assistants give you one model family and let you pick between a fast small version and a slow large one. RafayGen instead exposes six named agents — Fast, Everyday, Deep, Complex, Thinking and Frontier — and each one is wired to a different provider and model lane chosen for a kind of work.

The practical consequence is that the same prompt genuinely produces different work depending on which agent you ask, because a different model wrote it. Ask Fast to explain a concept and you get a crisp two-paragraph answer in under a second. Ask Thinking and you get a worked derivation that takes fifteen seconds and shows its steps. Neither is the better answer in the abstract; they are answers to different questions you happened to phrase the same way.

There is a second reason for the multi-provider design that has nothing to do with quality: single-provider assistants go down when their provider goes down. Routing across several means a rate limit or an outage at one lane degrades the service instead of ending it. The agent picker is the user-facing half of a failover system.

What it produces besides text

• Real-time spoken conversation through the voice orb, in English and Urdu, with the answer streaming as audio before the text has finished generating.
• Images, via FLUX running on hardware RafayGen operates directly — text-to-image, editing an uploaded picture by describing the change, restyling an existing image, and upscaling.
• Office files that are actually files: .pptx presentations, .docx documents, PDFs and .xlsx spreadsheets you download and open, not text you paste into Word yourself.
• Short generated video, from a text description or by animating an image.
• Urdu OCR — scanned Urdu books, photographs of pages and image-only PDFs turned into searchable, copyable text.

What it is not, and what it gets wrong

RafayGen does not train its own frontier models. It routes to providers that do. Anyone claiming a small independent team has built a competitive foundation model is either misusing the word or misleading you, and it is worth knowing which products are model builders and which are interfaces — RafayGen is firmly the second.

It inherits every failure mode of the models underneath it. It will state incorrect facts fluently. It will invent citations, statistics and quotations that look exactly like real ones. It is weakest precisely where it sounds most confident: specific numbers, named sources, recent events, and anything about a small local institution that the training data barely covers. The document engine checks that a generated deck has the structure it claims and that its links resolve; it cannot check whether the content is true, and no such checker exists.

Urdu output is markedly better than most global products but is not native-speaker quality, particularly for formal or literary registers. Roman Urdu comprehension is good and occasionally misreads an ambiguous spelling. Voice transcription degrades with background noise and heavy code-switching.

None of this is a reason not to use it. It is a reason to use it as a fast, tireless first draft and a thinking aid, and to keep a human in the loop for anything that matters.

How access and payment work

The free tier is permanent rather than a trial. It covers multi-agent chat in English, Urdu script and Roman Urdu, and you can hold your first conversation as a guest without creating an account at all — those guest conversations are deliberately ephemeral and are not attached to any identity.

The paid tiers cover the parts with a real per-request hardware cost: the voice orb, image and video generation, document generation, and higher limits. They are payable from Pakistan through NayaPay or a UBL bank transfer verified by hand, because the alternative — requiring an international card — would exclude most of the intended audience at the checkout page for reasons that have nothing to do with whether they want the product.

Nothing auto-renews, because there is no stored card to charge. When a month ends the account returns to free, and the history stays.

Who it is actually for

Students who need to understand a chapter tonight and cannot pay in dollars. Freelancers producing client deliverables who would rather generate a first-draft deck than build one from an empty slide. Small businesses writing product copy and customer replies in the mix of Urdu and English their customers actually use. Researchers and archivists with a shelf of Urdu material that exists only on paper.

If you already pay comfortably for a global assistant, work entirely in English, and need one specific frontier model, you are not the person this was built for and you should keep what you have. The honest pitch is narrower than the feature list: it is for people the other products lose at the payment page or the language barrier.