.Liveformerly AirVoice AI.
FinSense AI
A bank runs on unstructured Bangla: spoken down a phone line, handwritten onto a form. This turns both into data your core system can use.
95-99%
field extraction accuracy
.The problem.
Two queues with the same cause
Every retail bank has the same two queues. One is the call centre, where customers ask in Bangla for things the core banking system can only answer in structured queries. The other is the documentation desk, where account opening forms, KYC packets and loan files arrive as paper: printed, handwritten, photocopied, or photographed at an angle in bad light.
Both exist for the same reason. The bank's systems take structured input and its customers do not produce any. People have been doing that translation, which is slow and expensive, and inconsistent in exactly the way regulators dislike.
The usual fixes move the problem rather than closing it. An IVR menu makes callers navigate the bank's data model instead of asking a question. A template-based OCR works until the branch prints a new form.
The bank's systems take structured input, and its customers do not produce any.
.The approach.
Speech and documents are the same problem
They look like different problems and are usually bought as different products. Both come down to recovering structure from Bangla that was never written down in a structured form. Running them on one model stack is what makes the accuracy transfer, because the language understanding that reads a handwritten occupation field is the same understanding that parses a spoken request.
No template is registered in advance. The system reads the document rather than matching a layout, which is why it survives the photocopy of a photocopy that a real branch actually sends.
What FinSense is
One system. Two surfaces.
FinSense turns everything your customer says and everything your customer signs into data your core banking system can use. A bank runs on unstructured Bangla. Sometimes it arrives as a customer speaking, sometimes as a form somebody filled in by hand.
Front of house
Where the customer talks
- Conversational onboarding in place of forms
- Send money, pay bills and apply for products by speaking
- Runs inside the bank's own app, not beside it
Back of house
Where the paper becomes data
- Reads printed and handwritten Bangla and English
- Sorts, splits and pulls the fields out of any bank document
- Pushes clean data into the systems you already run
One Bangla speech and language stack under both surfaces. Built in house over eight years, not licensed from a vendor who has never seen a Bangla document.
Bangla speech accuracy
Languages supported
Or private cloud you control
Where does the data go?
The models are ours rather than an API we resell, so they run on premise in your data centre or in a private cloud you control. Every extracted field is traceable to the pixel region it came from, and every interaction is logged. For a regulated buyer that is usually the deciding factor.
Surface one
Where the customer talks
Voice AI Onboarding
The customer speaks and the account opens. Identity, document capture and KYC data collection happen in one conversation, in Bangla, without a form.
- 84%
- Completion rate
- 60%
- Faster KYC
Voice Navigation
Send money, pay a bill, check a balance, or apply for a loan by saying so. A multi screen journey becomes one sentence and an authentication step.
- 10s
- Per transaction
- 41
- Languages
Surface two
Where the paper becomes data
Four steps. It is exactly what your documentation desk does today, done by the system instead.
- 01
Sort by type
An incoming document is identified on its own. Account opening form, trade licence, national ID card, statement. Nobody configures a rule for it first.
- 02
Split the batch
A scanned stack of mixed paperwork is separated into individual records, using both how the pages look and what they say.
- 03
Pull the fields
Names, numbers, dates and amounts come out as clean structured data, from printed text and from handwriting.
- 04
Push to your systems
The finished record lands in core banking, in the document management system, and in the compliance trail. No re keying anywhere.
Built for Bangla bank documents specifically. Printed and handwritten, English mixed in, poor scans included. No layout is registered in advance and no template is maintained.
The hard case
A handwritten Bangla form, badly scanned
This is the one nobody else can read. It is also the one your branch network produces every day.
What the system is handed
- A photocopy of a photocopy, shot on a phone at an angle
- Ballpoint handwriting, Bangla and English on one line
- Fields written outside their boxes, corrections struck through
- A stamp and a signature laid across the printed text
- Bangla numerals in one field and English numerals in the next
What comes back, in seconds
- Applicant name
- National ID number
- Date of birth
- Permanent address
- Mobile number
- Occupation
- Account type
- Signature, matched to sample
Field level accuracy
Per document
Templates maintained
Extracted fields shown are a representative record, not a real customer.
Hand us a document we have never seen. We will run it at this table.
Start the forty five day pilot.
One document type or one onboarding journey. You pick the scope, we prove it on your own documents.
.Go deeper.
The full deck
Everything above, with the detail an evaluation actually needs.