Skip to content
GRAVITON AI · TWO AGENTS IN PRODUCTION

One agent fills the application. The other reads the whole file.

The Sales Agent sits with the field officer: it captures the customer’s documents, parses them and fills the entire loan application in minutes — in the customer’s language. The Credit Agent reads the finished file end to end, establishes the facts by arithmetic and writes the memo with its proof.

2 agents in production Minutes to a filled application Proof on every claim
GRAVITONCredit Agent 02 The file ●
ESTABLISHED BY ARITHMETIC
Monthly income overstated by 80% versus document
Form says ₹15,000; the bank statement says ₹3,000. The more authoritative source is the document.
Name on the bank statement is not the applicant
The file is in the name RAMESHBHAI K. PATEL; the statement carries KANTILAL S. PATEL.
SAVED REPORT · LA-2026-04417 · RUN 2026-08-20 15:43₹128 · THIS RUN
✓
SALES AGENT · हिन्दी
Application filled — 38 of 41 fields · 4 min 12 s

01 — THE SALES AGENT · IN THE FIELD

From a photograph to a filled application. In minutes.

The Sales Agent works where the loan starts — on the field officer’s phone, with the customer, in the customer’s language. Demo file, live layout.
SALES AGENT
LA-2026-04417 · LAP · ₹3.80 L
हिन्दी ▾
1 Capture ✓2 Parse ✓3 Fill ●
FROM THE OFFICER’S CAMERA
DOC Aadhaar · front + backName, DOB, address read READ ✓ 99%
DOC PANNumber verified against name READ ✓ 98%
DOC Bank statement · 1 dayAccount opened 28-10-2025 · one ₹3,000 credit READ ✓
DOC Property deed · 12 pagesSurvey no. 214/2 found READ ✓
DOC Customer signatureNeeded on 3 fields PENDING
FILLED INTO THE APPLICATION
Applicant nameRameshbhai K. Patel AADHAAR
Date of birth01-06-1998 AADHAAR
PANABCPP••••K PAN
Monthly income₹15,000 declared CUSTOMER
Property survey no.214/2 DEED
आधार और पैन पढ़ लिए। बैंक स्टेटमेंट सिर्फ़ एक दिन का है — 28-10-2025, ₹3,000 का एक क्रेडिट। मासिक आय कितनी लिखूँ?
₹15,000
ठीक है, ₹15,000 भर दिया — स्रोत: ग्राहक। 41 में से 38 फ़ील्ड पूरे, बाकी 3 के लिए हस्ताक्षर चाहिए।
APPLICATION · 38 OF 41 FIELDS4 MIN 12 S
Send to Graviton LOS →

EVERY FILLED FIELD CARRIES THE DOCUMENT IT CAME FROM

01
Captures from the officer’s phone

Aadhaar, PAN, voter ID, bank statements, salary slips, ITR, property papers, RC — photographed once at the doorstep, read once. No scanner, no branch visit to key it in.

02
Parses, then fills — not a separate form

Every field lands in the Graviton LOS application with its source document attached, so credit sees where a number came from and nothing is re-keyed downstream.

03
Speaks the customer’s language

Voice and text. The officer and the customer see the same conversation, and the agent asks for what is missing instead of leaving a blank.हिन्दीગુજરાતીमराठीதமிழ்తెలుగుಕನ್ನಡবাংলাEnglish

04
Minutes, not a second visit

The application is complete before the officer leaves. What is still missing is listed on the spot, not discovered a week later by the branch.

38of 41 fields filled 4documents read 1visit 4:12on the demo file

02 — THE CREDIT AGENT · ON THE DESK

The whole file, read. The facts, established by arithmetic.

Every document, ledger line and bureau pull on the file, read end to end — then the memo, each claim with the proof behind it. Switch views to see what the credit head gets. Demo file.
GRAVITONCredit Agent
ESTABLISHED BY ARITHMETIC6 FINDINGS · 4 RED · 2 AMBER
Monthly income overstated by 80% versus document
Form (CUSTOMER INFORMATION) says ₹15,000; document (bank_statement · statement_oct.jpg) says ₹3,000. The more authoritative source is the document.
Property value differs between two documents: ₹1,29,200 vs ₹9,52,674 (86%)
The technical valuation report states ₹1,29,200; the estimate states ₹9,52,674. These are separate documents, so this is a discrepancy on the file rather than a reading error. Establish which is current.
Banking shows ₹0 a month against the ₹25,000 the household declared
Total credits of ₹3,000, all of them on 28-10-2025, the day the account was opened. A single day of history cannot evidence a monthly income — either the money arrives somewhere else, or it is not there. Ask for the account the household actually operates on.
Name on the bank statement is not the applicant
The file is in the name RAMESHBHAI K. PATEL, but the statement (statement_2.jpg) carries KANTILAL S. PATEL. A document of this type should be in the applicant’s own name.
No existing obligations on the file, so FOIR has no denominator
FOIR was not computed. Two policy rules depend on it — #33 and #13 — and neither could be evaluated, which is not the same as either having passed. Supply the obligations and both can go to an approver.
Enquiries read as 2 from the bureau report, but the system holds 0
Both values come from machine sources, so this is most likely an extraction error rather than a discrepancy on the file. Verify against the bureau pull before acting on it.
Saved report · LA-2026-04417 · RAMESHBHAI K. PATEL · run 2026-08-20 15:43   Read-only — the live tool is not running. ₹128 · THIS RUN
READS · DOES NOT DECIDE
The credit head signs. The agent shows its working.
By arithmetic, not opinion

Income, LTV, FOIR and the gaps between documents are computed from what the file contains. A finding names the two sources and which one wins.

Proof on every claim

Each line in the memo carries the document, the field and the value it came from, so the reviewer checks the claim, not the agent.

Cost on every run

What the run cost is printed on the report — per file, in plain figures — so the desk knows exactly what a review is worth.

THE DESK VIEW

It also keeps the desk honest.

Every file the agent has read sits in one list with its memo and its risk review, complete and offline. Officers are scored on outcome against what their own book predicted — a desk view, not a case view.

CREDIT REVIEWRISK REVIEWDEVIATIONSOPENS OFFLINE
GRAVITONcredit & risk reviews
Case reviews

Each file below was read and written up beforehand. Opening one loads a complete, self-contained record — the memo, the reasoning and every document behind it.

Credit department — how the desk is performing
Officers scored on outcome against what their book predicted. Not a case file; not about one borrower.
SHOWEverythingCredit review (1)Risk review (3)3 of 3 files
RAMESHBHAI K. PATEL
LA-2026-04417   LAP_GUJ   Rs 3.80 L
No credit reviewRisk review1.0 MB · opens offline
Suresh Yadav
LA-2026-05486   LAP_RAJ   Rs 5.00 L
No credit reviewRisk review2.1 MB · opens offline
Vikram Meena
LA-2026-08301   LAP_RAJ   Rs 5.00 L
Credit review2.8 MB · opens offlineRisk review1.2 MB · opens offline
3 deviations1 unconfirmed#33 LTV+FOIR > 100 and ≤ 110, #59 Low development, #40 Any loan having > 60 DPD (except Gold) · 1 found in only one of three passes and not carried

03 — WHERE THEY SIT

Two agents, one file, one ledger.

Neither agent is a separate tab. The Sales Agent writes the LOS application; the Credit Agent reads it back with the ledger and the bureau; decisions execute through BRE and land in LMS and Gravisights.
FIELD
Officer’s phone
Documents photographed at the doorstep, in the customer’s language.
SALES AGENT
Capture · parse · fill
The application, complete, with a source on every field.
GRAVITON LOS
The file
One record — documents, ledger lines, bureau pulls, valuation.
CREDIT AGENT
Read · establish · write
Findings by arithmetic, deviations against policy, the memo with proof.
BRE → LMS · GRAVISIGHTS
Decision, then the book
The credit head decides; BRE records the path; the loan lives on in LMS and shows up in Gravisights.
Recommends, never decides

The Sales Agent fills; the officer confirms. The Credit Agent reads; the credit head signs. Every run says what it did not check.

Consent and an audit trail

Each capture is consented; each claim, field and run is logged with its source, replayable for an auditor.

Explainable through BRE

Anything that becomes a decision executes through Graviton BRE, so the rule path and version are stored alongside the memo.

Data stays in India

Documents, ledgers and memos are processed and stored in India, cloud or on-premise, on the same platform as the rest of Graviton.

FAQ

Graviton AI, in practice.

Two agents in production today. Ask us which document types and languages your book needs first.

Pick an agent. We'll run it on your files.

Ten of your own files through the Credit Agent, or one product through the Sales Agent in the field — the memo and the filled applications are the demo.

✓Response within one working day ✓Regulation-aligned: e-KYC, consent, audit trails, data in India ✓Every decision explainable through Graviton BRE

Book a demo

Thirty minutes with a product specialist, on data shaped like your book. We respond within one working day.

Request a demo

Or email neeraj@kugelblitz.xyz