Built for the Gnani Great Indian AI Internship Challenge. About the project
Voice service desk
for bank customers
Borrowers call, speak in any of 11 Indian languages, and finish the request on the same call. The agent never takes a password, PIN or OTP.
Caller
Maine is mahine ki EMI already pay kar di hai.
Agent
Thank you. I can see a payment of ₹12,450 received on 5 Nov against your loan ending 2046. I have raised a reconciliation ticket and paused reminders.
SMS · Sahyadri Finance Bank
Your account details were requested on a call. Full details: sent to your registered mobile ••••7310.
In numbers
11
Indian languages
23
gateway actions
4
security tiers
SMS only
full details go to the registered mobile
Tier 3
OTP, PIN and password never asked
Gnani
Prisma speech-to-text and Timbre voices
The problem
The call-back problem for lenders
Lenders automate outbound calls, but borrower call-backs are still handled by people. Borrowers wait on hold, repeat themselves, and have no safe way to get help.
RBI Ombudsman
Complaints keep rising while call-backs wait in a queue.
13.34 lakh
complaints to the RBI Ombudsman in FY 2024-25
29.25%
of them about loans and advances
70%+
of calls to RBI's contact centre are in Hindi
Source: RBI Ombudsman annual report, FY 2024-25.
Branch IFSC lookup
Public branch details, no sign in needed.
Masked account, customer ID, mobile and email
The agent only ever speaks the masked form.
Full details by SMS to the registered mobile
Never read aloud. Sent only to the number on file.
Password-reset guidance
Steps for the app or net banking. The secret is never handled.
EMI amount and due date
Read from the loan record after the caller is verified.
“I already paid” check with a reconciliation ticket
Date, amount and UTR are read back, matched to the ledger, and reminders pause.
EMI date change
Eligibility is checked, then the request is raised after a read-back.
Foreclosure quote and closure request
Outstanding amount and charges today, then a closure request.
Statements and interest certificates
Raised as a request and delivered to the registered channel.
Several requests without re-verifying
Verify once, then work through every request on the same call.
Mid-call language switching
English, Hindi, Hinglish or Marathi, whenever the caller changes.
Warm handover with a structured case record when unsure
A person gets the summary, what is done and what is pending.
Every spoken figure comes from the bank database, so the language model cannot invent a number.
Design rule of the gateway
See the security tiersHow it works
How a call runs
Pipeline
Every action passes the gateway. Nothing reaches the model unmasked.
- 01
Speak
The caller holds to talk. Gnani Prisma turns speech into text.
- 02
Verify
Registered mobile plus date of birth or loan last four digits.
- 03
Act
The agent calls one of 23 gateway actions. The gateway checks the tier, masks values and writes the audit log.
- 04
Confirm
Gnani Timbre speaks the reply. Sensitive details go by SMS to the registered phone, never read aloud.
Security
Security tiers written in code
Passwords, PINs, OTPs, CVV and card numbers are never asked, accepted, stored or returned.
- Masked outputs only
- Out-of-band delivery to the number on file
- Three failed verifications, then human handover
- 0
Tier 0 Public
No verification
Branch lookup, IFSC check, reset guidance
- 1
Tier 1 Verified
Registered mobile + date of birth or loan last 4
Masked values, EMI details, SMS delivery, hardship handover
- 2
Tier 2 Approved
Tier 1 + approval tapped in the bank app
Amounts, payments, foreclosure, every data change
- 3
Tier 3 Never
Cannot be unlockedPasswords, PINs, OTPs, CVV and card numbers are never asked, accepted, stored or returned
11 Indian languages on one call
Speech recognition and voices are Gnani Prisma and Timbre.
- EnglishEnglish
- हिन्दीHindi
- HinglishHinglish
- मराठीMarathi
- বাংলাBengalibeta
- ગુજરાતીGujaratibeta
- ಕನ್ನಡKannadabeta
- മലയാളംMalayalambeta
- ਪੰਜਾਬੀPunjabibeta
- தமிழ்Tamilbeta
- తెలుగుTelugubeta
- Callers can switch language mid-call
Measured results on one machine
1,000
concurrent simulated calls, 0 errors
94/100
scripted calls passed in the four core languages
139/140
scripted calls passed in the seven beta languages
Single-machine numbers: load generator, gateway, Postgres and Redis share one box. End-to-end voice latency is not measured at scale.
Two sign-in doors one for each side of the desk
Bank staff
- Bank manager: bank, ops, collections
- Engineer: dev lab, metrics, test cases
- Admin: everything
About LoanLoop
LoanLoop is an entry for the Gnani Great Indian AI Internship Challenge. Speech is Gnani: Prisma for recognition, Timbre for voices.
Is this a real bank?
No. Sahyadri Finance Bank is a fictional lender.
Will the agent ask for my OTP or PIN?
Never. If a caller says one aloud it is caught before the language model sees it and is not stored.
Which language model answers?
A local open model for English, Hindi, Hinglish and Marathi, and Sarvam for the other seven languages. Speech is Gnani.
What happens when the agent is unsure?
It hands over to a person with a structured case record.