AI agent development · From Udupi, across India
We build AI agents that do one named job inside your own systems, such as sorting enquiries, drafting replies or reading supplier bills, and then wait for a person to approve their work. Built and looked after from our studio in Udupi.
In short
An AI agent is software that does one business job in several steps: it reads a request, looks up the business's own records, such as stock or orders, and prepares a reply or an entry. Whirl Designs, a studio in Udupi, builds agents like this for businesses in coastal Karnataka and elsewhere in India, and each one waits for a person to approve its work before anything goes out.
Start here
Not every repeat job needs AI. Routine work can come off your team's desk in three ways, and we build the simplest one that does the job.
Best for: The same steps every time, no judgement needed
Best for: Questions answered from your own written content
Best for: Jobs that need reading, checking and a draft
When a fixed rule can do the job, we will tell you and build that instead, even if you came to us asking for AI.
Our work
No client of ours runs an AI agent yet, so this page has no client project to show. The approval screen at the top is a demo, and the cement order in it is made up. What we can point to is smaller and our own: the assistant on our website, which you can try.
It already has two of the parts an agent needs. It works only from this site's own writing, about 260 passages of service pages, FAQs and articles, and when a question matches a written FAQ closely, it gives that FAQ's answer unchanged instead of writing a new one. It also knows when to stop: a question it cannot answer is sent to a person on Telegram, whose reply lands back in the visitor's chat. What it does not do is act inside a business system. That part, with an approval step in front of it, is what a client agent adds.
Closest projects
Padmavathi Tours & Travels
Travel & transport · Udupi
A chauffeur-hire company with almost no online presence, rebuilt as a 190-page site that answers the exact searches its customers type — every vehicle, every route, every town.
Jayalaxmi Jewellers
Retail jewellery · Karkala
A brand site for a gold house trading since 1990, built to feel like the shop it represents — and backed by a full retail ERP handling billing, savings schemes and old-gold exchange.
ERP Sutra
Software products · Karnataka
Our own product arm — a shared platform layer with two industry ERPs built on top of it, both live with paying users rather than sitting in a pitch deck.
What you get
The AI model is only one part. The pieces around it decide whether the agent can be trusted with your customers. Every one of them is agreed in writing first.
How it runs
Each step ends with something your team can see and check before the next one starts.
Sorting enquiries, drafting replies to order requests, following up on enquiries that went quiet, or reading supplier invoices into purchase entries. We choose the one that takes the most staff hours and where a mistake would do the least harm, then watch the person who does it today.
A first version works through real enquiries or documents from the last few weeks, away from customers. You mark each draft right or wrong, and we fix what it gets wrong.
It gets read-only access to your stock, orders or billing. Anything it writes, such as a purchase entry from a supplier bill, stays a draft until a person saves it.
Every draft waits for a person. Your team works from the approval screen for a few weeks, and the log shows what the agent got right and what they changed.
Sorting, labelling and internal notes can later run on their own if the log shows they are reliable. Replies about orders, and anything that touches money, stay behind approval.
AI providers update and retire their models. When that happens, we run your test set again, fix what has changed and tell you before the new version goes live.
Not a fit
Questions
A chatbot answers questions; an agent does a job. An agent reads a request, looks something up in your own system, such as stock or an order, and prepares the next step, like a reply or a purchase entry. Because it can reach your records, how its access is set up matters more than which AI model it uses. Ours get the narrowest access the job needs and stop for a person before anything is sent.
Yes, AI models sometimes give a wrong answer in a confident voice, so we build every agent as if that will happen. Ours work only from your records and written content, give a written FAQ answer unchanged when a question matches it closely, and pass a case to a person when a record is missing instead of guessing. A person also checks every reply about an order before it goes out.
It can be, if what the agent sends and keeps is decided in writing before it is built. An AI model run by a provider such as Anthropic or OpenAI receives the text the agent passes to it, under that provider's terms, so we send only what the job needs and keep details like bank account numbers out where we can. We also agree what is kept and when it is deleted, because India's data protection law, the DPDP Act, covers the customer details an agent handles.
Yes. The agent connects through the WhatsApp Business API, the official route that lets software send and receive messages on a business number. Customer messages reach the agent, and its drafts go out once someone approves them. WhatsApp lets a business reply in its own words within 24 hours of the customer's last message; after that, and for messages your business sends first, such as reminders, it needs a template approved by Meta. We built a jeweller's retail system that sends receipts and reminders this way.
Yes. Current AI models read and write Kannada well enough for everyday replies, including messages typed in a mix of Kannada and English, or Kannada written in English letters. The agent answers in the language the customer used. Before launch, staff who speak Kannada read test drafts in both languages and mark any that sound stiff or wrong, and the approval step catches the rest.
A first agent for one job usually takes four to eight weeks. That covers writing the job down, testing on your past cases, connecting to your software and a few weeks of running with approval switched on. Older software with no way to connect can add time, and we will tell you that at the start rather than halfway through.
Often, yes. Tally can share data with other programs while it is open on your own computer, and many ERP and billing tools offer their own way in for other software. What your version and setup allow decides how the agent reads from it, so that is the first thing we look at. If the records live only in spreadsheets, we may suggest moving the key parts into a proper system first.
No. You need one person who knows the job well enough to check the agent's drafts in the first weeks, and nobody who writes code. We pick the AI model, usually Claude from Anthropic or a model from OpenAI, and look after the tests and connections. The AI account is opened in your business's name, so the account, the agent's instructions and its logs stay yours if you ever change developer.
Worth reading first
Clean, documented, tested code on a modern, proven stack — the foundation that keeps your ai agent development fast and dependable for years.
Ready when you are
Tell us the one job that takes up most of your team's day, and we will write down how an agent would handle it and where a person checks its work, or tell you honestly that a simpler tool will do.