← cojoy/live prototype

The agent that hands your credit team a decision-ready file.

It observes every source — bureau, banking, KYC, GST, documents — triangulates them, and questions the customer like a careful credit officer, in their own language, to assemble the file a fair assessment needs. It runs the bank's own policy as a rules engine to match the right product and never sends anyone to a dead-end — the credit decision stays with the bank. Built for reach and financial inclusion across India's tier-2/3 diversity, not just a better form.

Every conversation below is a recorded, deterministic replay of the real engine — identical every time, no live-LLM variance.

01The moat

It cross-examines — and it understands

The agent doesn't just collect a form. It reads the bureau and bank statements against what the customer said — surfacing the fuller story a form would miss, and questioning a figure that doesn't add up. The same skill serves the genuine customer and screens the fraudulent one. Switch between both below.

Try it live → /test
recorded · deterministic
Now, moving on to your income — what's your monthly take-home after all deductions?
My monthly take-home income is ₹75,000.
The agent · cross-examination
Understood — ₹75,000. I want to make sure I've got your income right: your bank statements show salary credits closer to ₹52,000 a month, a bit different from the ₹75,000 you mentioned. Could you help me understand which reflects your actual take-home?
My actual take-home is ₹52,000. My salary account is with HDFC Bank.
Thank you for clarifying — ₹52,000, salary account with HDFC. Next, what are your existing total monthly EMI obligations across all loans and cards?
₹15,000.

Stated ₹75,000 · bank statements observed ₹52,000 · the agent raised it in its own advisor voice, the applicant corrected down — a 31% income drift, recorded as a fraud-screening signal.

02The artifact

Every conversation becomes a decision-ready file

At close, the lender's risk team gets a file: the cross-source discrepancies the agent flagged, the file checked against the bank's own credit policy (version-stamped for audit), and the best-fit product across the shelf. The credit call stays with the bank.

Fraud-screening signal · 1 prior-answer conflict
Monthly take-home income (INR): ₹75,000 ₹52,000 · 31% drift (75000 → 52000)
NBFC personal loan — India
checked against policy v1.0.0
meets policy
Min credit score
≥ 700 · observed 728
meets
Current obligations (FOIR)
≤ 55% · observed 29%
meets
Min monthly income
≥ ₹25,000 · observed ₹52,000 (observed)
meets
Age band
21–60 · observed 31
meets
Evidence: bureau
observed · observed observed
meets
Evidence: banking
observed · observed observed
meets
Evidence: kyc
observed · observed observed
meets

The file clears the cut-offs — but the income drift above rides along on the record, so your credit team decides with eyes open. Thin-file / new-to-credit applicants may still be considered with stronger income proof or a co-applicant — do not hard-decline on a missing score alone; surface the path. FOIR may be relaxed toward 0.60 for salaried applicants above ₹1L take-home with a clean bureau and stable tenure.

03Financial inclusion, as an engine

Nobody is turned away

A customer who fails the product they applied for isn't rejected — the engine ranks them across the whole shelf and routes them to one they DO qualify for: a new-to-credit loan, a secured card, a gold loan.

Asha Devi · bureau score 642 · income ₹28,000 below the standard personal-loan bar. The engine ranks the whole shelf and finds where she fits:
Personal loan — new-to-credit · thin-file
↳ routed here — best fit
qualifies
Secured credit card · secured
qualifies
Gold loan · secured
needs info
NBFC personal loan — India · standard
disqualified
Personal loan — relaxed-FOIR experiment · experiment-A
disqualified
Credit card — India
disqualified
MSME business credit — India
disqualified

No dead-end: a customer the standard product declines is routed to Personal loan — new-to-credit, which she qualifies for — the financial-inclusion thesis, as an engine.

And it goes further

04 · No wasted questions

It knows what they want

A customer who already knows they want a gold loan answers ~9 questions — the gold + KYC it actually needs — not the 21-field personal-loan interrogation. The questions adapt to the product.

Try the gold loan → /test
05 · A relationship, not a funnel

The 11th interaction is cheaper than the 1st

Months later, the next product is a 90-second start: the agent remembers the household, employer, and bureau identity, and only asks the delta — never a re-onboarding.

06 · Policy as config

Train it on your policy, not an IT project

Paste a credit-policy document and it compiles to a structured, validated policy that applies in the next conversation. Edit a cut-off — score, FOIR, income floor — and the next file is re-checked against it, with no deploy.

Open the catalog console → /admin/catalog