Get your AI agents working in production.
MedhAI is an AI engineering firm for banks, lenders and payments businesses. We work with your team to build AI agents on your own systems and get them into production.
We've done it before. At a large payments company, Atul led the program that put voice and chat agents into production. They handled tens of thousands of customer conversations a day.
What we have built and run.
Deployed · client work
Voice and chat agents for a large payments company
Voice and chat agents in customer service, sales, retention and merchant support, handling tens of thousands of conversations a day. Atul led the program. The client's teams built and ran the individual agents.
Deployed · client work
An in-house model gateway
Atul led the replacement of a contracted model gateway at the same company. Ten use cases from eight teams moved onto it in the first month, with cost-based model routing and one view of spend.
Demonstration · sample data
A failure drill that caught a duplicate case
We made a case system time out just after it saved a case. The agent retried and opened the same case twice, because the request carried no idempotency key. We recorded it and listed the fix. See the drill
Demonstration · sample data
Collections calls in English, Hindi and Punjabi
Built for Canadian callers who would rather speak their own language. The agent checks who is calling before it mentions a balance. In one recorded Punjabi call it tried to log "today" as the payment date. The caller never said that, so the action was blocked. See the trace
We take one or two engagements at a time, and Atul works on each one. Right now we're looking for first pilots with Canadian lenders and payments businesses.
Workflows we build.
Each one is a use case where an agent automates part of the work, then checks it actually got done.
"Where is my payment?"
Find the transaction's actual state across systems, explain the next step and open an investigation when it's needed, without telling anyone the money arrived before it did.
Account changes that need approval
Verify the customer, check risk signals, then make the change, hold it or send it for a second person's approval, and confirm the resulting record.
Payment commitments
Discuss a balance, record a promise only with the date the customer gives, and hand the case to a person when the rules require it.
Payroll and settlement exceptions
Gather the records behind a failed or mismatched payment, prepare a correction for approval and route what stays unresolved.
Merchant support and disputes
Answer merchant questions from their own records, gather the evidence for a dispute or chargeback, and hand the decision to a person.
The second and third are in the demonstration, recorded with sample data.
How much should the agent be allowed to do?
Most agents start by answering questions. Some prepare an action and wait for a person to approve it. A few act on their own, only inside limits you set, with a record of what the customer agreed to.
Answer
Look things up and explain them. Our merchant support chat work sits here.
Prepare for approval
Draft the correction or the change, then wait. In the demonstration, replacing an external account waits for a second person's sign-off.
Act within limits
Act, then confirm the record changed. In one test, an agent tried to log a payment date the caller never gave. It was blocked.
Rules in code decide what the agent may do. The model handles the conversation. We help you decide which steps belong at which level. Then we build them and test them against the failures your systems actually produce.
Most engagements start with a workflow review, then a pilot on one workflow, then more once it works. We agree up front how to measure it: tasks completed, errors, work sent back to people, cost per task. You keep the code and the test sets.
See what is ready and what needs work.
MedhWorks is where we write down what the agent may do, then record every test against it, including the ones that fail.
Atul Shridhar.
MedhAI is led by Atul Shridhar, who has spent twenty years building payments, data and AI systems in regulated markets. At Paytm he was founding architect of the payments bank's integration layer and built the company-wide data platform. At Paytm Labs in Toronto he was VP Engineering for the growth and personalization platform (50M monthly transacting users). Most recently he led AI platforms and agent adoption at a large payments company.
Before we start.
How do you handle data and access?
Our agents use your existing APIs and never read your databases directly. Before we build, we agree with your team where it runs, what access it needs and how data is handled. We don't need customer records for the first conversation.
What happens in the first conversation?
We talk through the workflow, the systems involved and what's stuck. Then we tell you whether an engagement makes sense.
Start with one workflow.
Tell us what the agent needs to do, which systems it must use and what is holding the deployment back.