KreditBeeDvar

Collections copilot

Every collections call, read and filed automatically

After each call, by the voice agent or by a person, a small model fills in the disposition: promise date and amount, reason, hardship, a request to stop calls. Code then decides the next step. The borrower's protections (stop requests, hardship, someone else on the line) come first.

Internal benchmark on synthetic data · run run-20261022172936-test · 2026-10-22 · 3 models, same instruction

C073 · Kannada · human agent call · 2026-10-19 17:37

Arjun Naik, Belagavi · loan PL46662655 · EMI ₹5,120 due 2026-10-10 (9 days past due)

  1. Agentನಮಸ್ಕಾರ, ನಿಮ್ಮ ಪರ್ಸನಲ್ ಲೋನ್ ಖಾತೆ ಬಗ್ಗೆ ಕಾಲ್ ಮಾಡ್ತಿದ್ದೀನಿ. ಈ ಕಾಲ್ ರೆಕಾರ್ಡ್ ಆಗ್ತಿದೆ. ನಾನು Arjun ಅವರ ಜೊತೆ ಮಾತಾಡ್ತಿದ್ದೀನಾ?

  2. Customerಹೌದು, ನಾನೇ ಮಾತಾಡ್ತಿರೋದು.

  3. Agentನಿಮ್ಮ ₹5,120 EMI ಕಟ್ಟುವ ದಿನಾಂಕ 10 ಅಕ್ಟೋಬರ್ ಆಗಿತ್ತು, ಇನ್ನೂ ಬಾಕಿ ಇದೆ. ಯಾವಾಗ ಕಟ್ಟುತ್ತೀರಿ?

  4. Customerನೋಡಿ, ಕಳೆದ ತಿಂಗಳು ಕೆಲಸ ಹೋಯ್ತು, ಆದರೂ 27ನೇ ತಾರೀಖು ₹2,600 ಹೇಗಾದರೂ ಕಟ್ಟುತ್ತೀನಿ.

  5. Agentಸರಿ, 27ನೇ ತಾರೀಖು ₹2,600. ಧನ್ಯವಾದಗಳು.

Disposition, as the model recorded it

Outcome
ptp
Promise date
2026-10-27
Promise amount
₹2,600
Reason
job loss
Hardship
Yes
Wants a person
No
Stop calling
No
Call back at
—
Next action (code): Hardship desk within 2 working days: restructuring or a payment holiday. Keep the promise of 2026-10-27 but pause late fees.

Behind the answer

Model
dvar:glm-5.3-flash
Time
1.76 s
Cost per call
$0.00027
Scorer
Every field right

Run run-20261022172936-test, 2026-10-22. Same instruction for every model, temperature 0. Cost is from tokens and list prices.

All 120 test calls, every model

Same instruction, temperature 0, one deterministic scorer. Critical failures are counted by call; one is enough to keep a model off that call type.

ModelEvery field rightCalls with a critical failurePromise dateHardshipp50 / p95Cost per 1,000 calls
Dvar small modeldvar:glm-5.3-flash96.7% (116/120)0100.0%100.0%2.02 s / 3.89 s$0.26
Low-cost frontieropenai/gpt-6-luna98.3% (118/120)1wrong ptp date 199.2%100.0%600 ms / 839 ms$0.11
Mid-tier frontieranthropic/claude-sonnet-5.590.0% (108/120)0100.0%100.0%534 ms / 694 ms$3.34

Tokens per call average 750 in and 360 out. Run run-20261022172936-test.

Written for the demo: every borrower, call and transcript (Hinglish, English, Hindi and Kannada), and the answer key. Computed here: every score, next action, latency and cost figure, from the frozen run. On a placeholder run the model replies are the answer key with seeded errors.