Three twins. Trained, tested, governed.
A twin of your best human agent, a twin of your customer base, and a twin of the journey itself — so a voice agent can learn from your floor, be tested against your callers, and rehearse a change before it ever speaks to a real customer.
A twin is not a clone of a person. It is a model — and a model is trained, tested and governed, or it is not allowed to speak.
Agent. Customers. Journey. Governed.
The twin of your best human agent.
Speech and reasoning fine-tuned on their consented, redacted calls — their phrasing, their pace, how they handle "salary next month". Trained inside your tenant, never pooled.
The twin of your callers.
The simulator that plays 1,240 scenarios drawn from your real call distributions — barge-in, already paid, wrong number, legal threat — in Hindi, Hinglish and Tamil. It is what the evaluation models test against.
The twin of the flow itself.
The stage graph of a collections or renewal journey, with its policies attached. Change a PTP window or a tone rule and rehearse it against the customer twin before it goes live.
Governed like everything else.
Every twin has a training set your reviewer approved, a scoreboard, frozen lines checked verbatim, a governance policy, a named approver and one-click rollback.
Your top collector's moves, at every desk.
One agent on your floor keeps promises when others get hang-ups. Their calls — consented, Aadhaar and PAN redacted — become the fine-tuning set for a twin that speaks the way they do and reasons the way they do. The twin inherits the move, not the memory: every amount and date it says is re-grounded against the ledger on that turn, and the frozen lines stay frozen.
The twin lives in your tenant. Its weights are yours, and it improves in the same weekly cycle as everything else.
Illustrative pair. The twin learns the move, not the transcript; every amount it speaks is re-grounded per turn. replace with a consented pair from a pilot
A simulator that plays your callers.
Every release is tested against a model of your customer base before it speaks — the same simulator the evaluation models use. Scenarios are drawn from the real distribution of your calls, so the twin gets angry as often as your callers do, in the languages they use.
| Scenario | Language | The customer twin plays | Outcome on v42 |
|---|---|---|---|
| already_paid | Hindi | "Kal UPI se kar diya" — gives a UTR only if asked twice | ✓ UTR looked up · closed · no repeat |
| salary_delay | Hinglish | Pushes payment to "next month", gets irritated on the second ask | ✓ PTP on salary date · token offered |
| barge_in_amount | Tamil | Interrupts mid-amount, asks "evvalavu?" again | ✓ restated once · no loop |
| wrong_number | Marathi | "He Sunil naahi" — refuses to say who they are | ✓ no account detail spoken · DNC noted |
| legal_threat | Hindi | Mentions a lawyer, raises voice, asks for a name | ✓ warm transfer · brief to human |
| outside_hours | Hinglish | Answers at 20:40 and asks to continue | ✓ RBI FPC hours · callback scheduled |
| branch_or_link | Hindi | Says "link" then asks "branch se ho jayega?" | ✗ asked twice · blocks release |
7 of 1,240 scenarios. Distributions drawn from the tenant's own call outcomes, refreshed weekly. illustrative · replace with a live scoreboard
A scripted test set only checks what you thought of. The customer twin is refreshed each week from graded production calls — new phrasings of "already paid", new reasons for delay — so the release is tested against last week's customers, not last year's assumptions.
Rehearse the policy change before it goes live.
Illustrative rehearsal. replace with a live run
The stage graph is the twin.
A collections journey is a graph: greeting, verify, classify, capture promise, dispute, close, transfer — with a governance policy bound to each stage. The journey twin runs that graph against the customer twin. Widen the promise-to-pay window, change the tone bucket for DPD 31–60, add a callback stage: you see promise, kept-promise, callback and transfer rates move before a single real call carries the change.
Nothing ships from a rehearsal.
A rehearsal is a report, not a release. The change still goes through the scoreboard, the frozen-line check, the named approver and the canary. What the twin gives you is a number to argue with in the room — instead of a week of live calls to find out.
Illustrative week. confirm cadence with the eval team
A twin is a release like any other.
The word "twin" earns no exemptions. The best-agent twin is a fine-tuned model with a curated, redacted, approved training set. The customer twin is an evaluation model with its own human calibration. The journey twin is a policy graph under governance. All three carry a scoreboard, a frozen-line check, a named approver and a rollback — and all three are deployed inside your walls.