Case study · Banking · GenAI
How a leading Indian bank put 4,000+ policies in front of 10,000+ employees with ShellKode FDEs.
Leading Indian bank · Banking
Staff used to dig through folders for the right circular. Now they ask, and eight agents find the answer and show its source.
10,000+
Employees on one secure GenAI assistant
85%
Answer accuracy on the bank's own policy questions
4%
Hallucination rate, with up to 6 cited sources per answer
8 weeks
Kickoff to production go-live
The challenge
A bank's rulebook, locked in a file system nobody could search.
More than 10,000 employees relied on 4,000+ policies, circulars and guidelines to do their jobs. All of it sat in an internal wiki that was really a plain file system, with no intelligent search. Finding the right rule meant clicking through folders and reading documents end to end.
Putting a chatbot on top of the files wouldn't have been enough. In a regulated bank, every answer has to cite its source, every document has to be visible only to the right roles, and every action has to be logged for the security team. That needed engineers working inside the bank's IT and information security teams, making decisions with them every day.
Slow file system
The internal wiki had no GenAI, so every answer meant opening folders and reading documents.
Answers had to be trusted
Staff needed to see the source document before acting on an answer.
Access by role
Documents had to be limited by role and category, not open to everyone.
Regulated from day one
Strong sign-in, a full audit trail, and a regulatory security review were required before launch.
What we delivered
Eight agents to answer from the bank's documents, and the controls to run them safely.
ShellKode FDEs embedded with the bank's IT, information security, and business teams for eight weeks, from design to go-live, to build the assistant on AWS in an Indian region.
A pipeline of agents
Strands agents on Amazon EC2 check each question, route it to the right domain, retrieve from Amazon Bedrock Knowledge Bases, and answer with Claude Sonnet, with citations.
Guarded and private
Amazon Bedrock Guardrails screen questions and answers, and calls to Bedrock stay on private VPC endpoints.
Only verified employees
Employees sign in using the bank's on-premises Active Directory over a Site-to-Site VPN, with a mobile OTP on every login.
Every action on record
User activity and every AI answer go to an audit trail in Amazon RDS and to the bank's security monitoring system in real time.
The outcome
Every policy answer now comes with its source.
85% accurate on real policy questions
Scored by the application team against known answers, with 100% on KYC and documentation.
4% hallucination rate
Retrieval keeps only strong matches and every answer cites up to 6 source documents.
2.3-second answers
For 10,000+ employees, across 4,000+ documents.
Live in 8 weeks
From kickoff on 3 August to production on 25 September 2026, with all RBI VAPT findings closed.
The stack
What it runs on.
- Agents and models
- Strands Agents on Amazon EC2, Claude Sonnet on Amazon Bedrock
- Knowledge and guardrails
- Amazon Bedrock Knowledge Bases, Amazon Bedrock Guardrails
- Access and network
- On-premises Active Directory, Site-to-Site VPN, mobile OTP, private VPC endpoints
- Audit and security monitoring
- Amazon RDS, the bank's security monitoring system
What's next
The platform is live. Next comes sharpening what it answers.
A new document category now takes an upload, not a new project.
- 01
Retiring superseded circulars, so charges and fees answers always come from the latest version
- 02
Answer feedback from staff to keep improving retrieval
- 03
Usage and answer-quality dashboards for business and compliance teams
Talk to us
Bring one document set. Leave with a secure, cited GenAI assistant plan.
30 minutes on one knowledge workload, mapped to the AWS architecture and controls behind it. No slides.