Live in production. Routing every message right now.
One AI. A different model every message.
Swaylo AI reads a message before it answers it, then sends it to the cheapest model that can actually do that job. Small and fast for text. One that can hear for a voice note. One that can see for a photo. It switches in the middle of a conversation, and the customer never notices.
The whole idea
Most messages don't need a big model. So most messages don't get one.
A single large model can do everything, and bills you for everything on every "hi". Swaylo AI classifies each message first and spends only what that message costs.
How it works
Four decisions, taken on every single message.
Not configured once and left alone. It runs again from the top each time somebody says something.
It classifies
Before anything answers, it works out what kind of message this is: words, speech, or a picture.
It routes
To the cheapest model that can do that particular job. Vision and speech are paid for only when they are actually used.
It speaks as one
Three models can touch one conversation and it still sounds like one assistant, with one memory and one tone.
It survives
If a provider stops answering it switches again, to the other one's cheapest model, without dropping the thread.
What the engine gives you
Cheap by construction, not by settling for a worse model.
The saving comes from asking the expensive models less often, not from asking them for less.
It hears
Voice notes are transcribed, in Indian languages, including an address read out loud into the phone.
It sees
Photos are read and described, and handwriting in them is typed out, so a scrawled address becomes usable text.
Ten Indian languages
It answers in the language it was written to, in that language's own script, or in Latin letters if that's how they typed.
Two providers, one voice
It is not tied to one vendor. If one stops answering, the conversation carries on somewhere else.
It degrades, it doesn't stop
The worst outage produces a cheaper answer, not an unanswered customer.
No keyword lists
No menu tree, no regex to tune, no "press 1 for". It judges the message; that is the whole mechanism.
Where it runs
Swaylo AI doesn't chat on WhatsApp. Shopdesk Assistant does.
The engine has no opinion about where the message came from. Selling on WhatsApp is a job built on top of it, the first one, not the only one it can do.
Shopdesk Assistant
The WhatsApp sales assistant for a small Indian online store, running on Swaylo AI. It handles every conversation the ads bring in, unattended, all day.
- Opens on the product they tapped in the ad, at its real price
- Quotes from a live catalogue instead of inventing a number
- Verifies the pincode against India Post mid-chat
- Books the order, and cancels it if they change their mind
- Handles Cash on Delivery and UPI advances
- Hands the chat to a human the moment it should
Anywhere else a message arrives.
None of the routing above is about WhatsApp, or about selling. It is about spending the least that will answer this message properly, which is true of any application with people typing at it.
- One engine, many front doors
- Built and run by one person, not a platform
Akshay C P founded and built both of them: Swaylo AI, and Shopdesk Assistant on top of it. One person in Kerala. The routing, the prompts, the ordering logic, the ads that bring the customers and the phone number at the bottom of this page. If something here is wrong, there is nobody else to blame for it.
Don't take my word for it
Message +91 99110 70741. That's Shopdesk Assistant, the one application plugged in today. Type at it, then send a voice note, then a photo. Three models will answer you and it will sound like one.
Start a chat