Behind the build
Virtual Relationship Manager
A relationship manager inside the banking app, answering in plain language.
AI · Model development · Fintech
Banks and fintech companies wanted their customers to get real answers inside the app, not a menu of options or a reply that sends them to a call center. We built a virtual relationship manager: two fine-tuned models that understand any banking question, typed or spoken, fetch the answer through the bank's existing APIs, and reply in plain language. It runs on the bank's own servers, and it is in use with fintech companies from banks to startups.
- Built for
- Banks and fintech startups
- Our role
- Model fine-tuning, core-system API layer, on-premise deployment
- Deployment
- On the bank's own servers
- Input
- Text and voice
- Response time
- Under 100 ms
More than a chatbot
Customers don't speak in menus. They ask in their own words, and the answer usually lives in the bank's core system.
A chatbot can reply, but it can't look anything up. This had to understand any banking question, typed or spoken, fetch the real data behind it, and answer the way a good relationship manager would, inside the banking app the customer already uses.
How it works
Every question goes through three steps.
- 01
Understand
A fine-tuned classification model reads the question, typed or spoken, and works out what the customer is asking.
- 02
Act
The core-system API layer calls the right service in the bank's core system and fetches the real data.
- 03
Answer
A fine-tuned text generation model turns the result into a clear reply in natural language.
Two models, one job each
We split the work between two models, each fine-tuned on banking customer queries, so understanding stays fast and every answer is written from the bank's real data.
Intent classification
An encoder model of about 166 million parameters. Small and fast, it decides what each question needs.
Response generation
A decoder model of about 3 billion parameters. It writes the answer in natural language from the data the core system returns.
Core-system API layer
The bridge between the models and the bank: it maps each intent to the right call and returns only what the answer needs.
On-premise deployment
Hosted on the bank's own servers, not ours, so the models and the customer data stay inside the bank's infrastructure.
Built for fintech
Every part of the system is specific to banking, from the queries the models were fine-tuned on to the core systems it connects to. It plugs into the APIs a bank already has, so nothing in the core system has to change.
That is what lets the same system serve very different fintech companies, from banks to fintech startups, each on its own infrastructure.
Where it stands
₹20MIn fixed deposits booked through the virtual relationship manager in a single month.
The virtual relationship manager is in use with multiple fintech companies, inside their banking apps, on their own servers.
Their customers ask in their own words and get an answer from real data, without waiting for a person or finding their way through the app.
1M+
Customers supported by the virtual relationship manager.
<100 ms
From a customer's question to an answer in their own words.
500+
Loan requests within fifteen days.
15%
Fewer calls to the call center.