(RAFAYGEN_AI)
Sign up free
RafayGen AI

One assistant.
Chat, voice, images
and documents.

Free to try, no card. Six chat agents on six different engines, a voice you can interrupt, images generated on our own GPU, and real office files — built and operated from Sindh, Pakistan.

RafayGen's voice orb — the live WebGL shader that visualises speech in the app
The voice orb, captured from the running app. Not an illustration — this is the shader you talk to.
Why it exists

Most AI products assume an international credit card, an English keyboard, and a user who treats Urdu as something to translate out of.

Every one of those assumptions is a wall, and in Pakistan you hit all three in the first ten minutes. RafayGen was built from the other side of them: local payment rails first, Urdu and Roman Urdu as inputs the engine actually understands, and a first conversation that works before there is an account to bill.

It is one product, not a suite. Chat and voice share an account, an engine and a memory of who you are, so a feature lands on both or it is not finished.

Chat

Six agents, six different engines.

Fast, Everyday, Deep, Complex, Thinking and Frontier each run on a distinct provider and model lane. The same prompt genuinely produces different work depending on which you pick — that is the point, not a side effect.

Routing across providers also means one upstream outage degrades a corner of the product instead of taking it down. The cost is real: six lanes are six sets of quirks, and most of the engine's defensive code exists because of that trade.

Abstract render: a dark indigo field of faint connected nodes, like a mesh of routes
Voice

A voice you can interrupt.

Speech in, a low-latency answer, and speech streaming back out while it is still being written — so you can cut in mid-sentence the way you would with a person. English and Urdu.

It is the same assistant as the chat window, with the same memory. Start a thought by typing and finish it out loud.

Abstract render: twisting violet ribbons coiling around each other
Images

Generated on our own GPU.

FLUX text-to-image, editing, img2img and upscaling — no separate tool and no separate tab. Ask in chat, then refine in replies: “same scene at night”, “remove the car”.

Iterating in a reply keeps everything the first attempt already got right, which is the part a fresh prompt throws away.

Abstract render: a dense field of faceted indigo crystal shards catching light
Documents

Files that actually open.

Decks, reports, letters and spreadsheets come back as real .pptx, .docx, .pdf and .xlsx, structurally validated before you are told they exist.

A download link that does not resolve never reaches you. That is enforced in code, not promised in a prompt.

Abstract render: a grid of stacked amber-lit blocks receding into shadow
Language

Type the way you think.

English, Urdu script, Roman Urdu — including a sentence that switches halfway through, which is how most people here actually write.

The Urdu OCR pipeline came out of the same work: scanned Nastaliq pages become searchable text and rebuilt PDFs. It was written to digitise Urdu literature, which is harder than digitising invoices, and it shows.

Abstract render: folds of dark violet silk falling in smooth curves
How an answer is chosen

Many possible answers.
One that arrives.

A language model does not look up a reply. It holds a distribution over what could come next and draws one sample from it — which is why the same question can be answered two good ways, and why one lucky answer proves nothing about a model.

That is the whole argument behind how we test this product: run it enough times to see the spread, then publish the number of runs next to the result. A score without a sample size is a claim, not a measurement.

Evidence

We publish a benchmark our own product came fourth of four in.

The open Urdu benchmark grades every model with the same code — no AI judge, no rubric to argue with — publishes the raw answers, and ranks RafayGen next to the systems it competes with. In the first complete run it placed last. The result went up unchanged, and the defects it exposed were fixed the same week.

Read the leaderboard and the method, or see the live numbers straight from the running system.

Pricing

Priced for where it is used.

Free
No card, no trial clock.The first conversation works as a guest, before an account exists.
Go — $15
Higher limits for regular use.
Plus — $50
For daily, heavier work.
Extreme — $100
The ceiling raised as far as it goes.
How you pay
NayaPay or UBL bank transfer, verified by hand.No international card required — that wall is the reason this exists.

The full comparison, including where RafayGen is genuinely worse than the alternatives, is in a ChatGPT alternative for Pakistan.

Open it and ask it something hard.

Guest access needs no account and no card. If it disappoints you, the benchmark probably already says why.