Build it, test it, run it
One dashboard for the whole life of an agent. These are screenshots of it, on a test workspace.
Start from a template or your own documents, then shape the call step by step.
- The agent builder reads your SOPs and FAQs, suggests a template and fills the draft in, round by round.
- The graph builder lays the call out as steps, each with its own instructions and exits.
- A safety transition is checked from every step, ahead of the step's own exits.
Find the broken path before a caller does.
- Scripted callers walk the graph in Hindi, English and Hinglish; each scenario passes or fails.
- Every run shows the path taken and the model's reason for each step, plus steps never reached.
- Simulated calls play the whole call through the voice pipeline, with recordings to download.
See how your agents are doing, and what needs you, on one page.
- Calls, completion rate, safety-step triggers and spend for the last seven days.
- Alerts for failing tests, a low wallet, calls ending in error and pipelines that are not ready.
- Simulated calls are kept out of the numbers unless you switch them in.
A payment reminder, in Hindi, start to finish
The caller switches between Hindi and Hinglish, confirms who they are on the keypad and promises a part payment. Then they report a charge they never made: the safety step takes over and hands the call to the fraud desk.
- Greet
- Keypad check
- State the dues
- Promise to pay
- Fraud desk(from any step)
- AgentAI disclosure said
नमस्ते, मैं सूर्या बैंक की एआई सहायक मीरा बोल रही हूँ। यह कॉल रिकॉर्ड हो सकती है। क्या मेरी बात राहुल जी से हो रही है?
Hello, this is Meera, an AI assistant of Surya Bank. This call may be recorded. Am I speaking with Rahul?
- Caller
Haan, main Rahul bol raha hoon.
Yes, this is Rahul speaking.
- Agent
धन्यवाद। पहचान के लिए कृपया अपने कार्ड के आख़िरी चार अंक फ़ोन के कीपैड पर दबाइए।
Thank you. To confirm it is you, please type the last four digits of your card on the phone keypad.
- Callerkeypad: •••• · identity confirmed
(types four digits on the keypad)
The digits go straight to the bank's check; they are never spoken, shown or recorded.
- Agentget_dues → ₹8,400 total · ₹1,200 minimum · due the 18th
पहचान हो गई। आपके कार्ड पर कुल बकाया आठ हज़ार चार सौ रुपये है, न्यूनतम देय बारह सौ रुपये, और आख़िरी तारीख़ अठारह है। आप कब तक भुगतान कर पाएँगे?
Thank you, that matches. The total due on your card is ₹8,400, the minimum due is ₹1,200, and the due date is the 18th. By when can you pay?
- Caller
Abhi poora nahi ho payega. Chaar hazaar ikkees tareekh tak de dunga, baaki agle mahine.
I can't pay all of it now. I'll pay four thousand by the 21st, the rest next month.
- Agentsave_promise → ₹4,000 by the 21st
ठीक है, मैंने दर्ज कर लिया: इक्कीस तारीख़ तक चार हज़ार रुपये। आप UPI, नेट बैंकिंग, ऐप या किसी भी शाखा से भुगतान कर सकते हैं।
Alright, I have noted it: ₹4,000 by the 21st. You can pay by UPI, net banking, the app or at any branch.
- CallerSafety transition: suspected fraud, from the promise step
एक मिनट… इस बिल में तीन हज़ार की एक ख़रीद है जो मैंने की ही नहीं। किसी ने मेरा कार्ड बिना पूछे इस्तेमाल किया है!
One minute… there is a purchase of three thousand on this bill that I never made. Someone used my card without my permission!
- Agenttransfer_to_human → fraud desk, with a handover note
आपने बताकर सही किया। मैं आपको अभी बैंक की धोखाधड़ी टीम से जोड़ रही हूँ। कृपया किसी को भी OTP, PIN या पूरा कार्ड नंबर न बताएँ।
You were right to tell us. I am connecting you to the bank's fraud team now. Please do not share an OTP, PIN or your full card number with anyone.
- Caller
Theek hai, please jaldi.
Okay, please be quick.
A scripted sample, not a recording: the lines follow the credit-card payment reminder template with the fraud safety transition the agent builder suggests for card agents; the bank, the customer and the amounts are invented.
Measured, not claimed
On our own benchmarks. The callers are synthetic test callers played as audio through our engine and our self-hosted models; none of these are customer calls. Each series names the model it ran on.
Gemma 4 26B-A4B on A100, 2026-10-20, 80-call card-servicing run
Card-servicing calls passed
80 scripted card-holder calls to an invented bank, through our graph agent and the real speech models. A call passes only when every expectation holds.
61of 80
- 0 verification bypasses in 80 calls
- No account step or bank operation ran for a caller who had not been verified.
- Fraud: 11 of 11 reached the fraud step
- Every caller who reported fraud was moved to the fraud step, from any point in the call; 9 of those 11 calls passed every other expectation too. No false alarms in the other 69 calls.
- 1.08 s from caller to first audio
- Median from the caller stopping to the agent's first sound at the caller's end, waiting for the end of the turn included; 1.40 s at the 90th percentile, over 253 spoken turns of this run.
- 526 ms of model time per turn
- Median time for the model to pick the next step and finish its reply, on a text replay of the same calls without speech.
What is still weak: English callers go through a Hindi transcriber today and pass less often (15 of 26), and abusive callers are handled well in only 2 of 5 calls.
Measured on Gemma 4 E4B, October 2026; 26B results to follow
Interruptions, before and after one round of engine work
Interruptions caught
Test callers talk over the agent mid-sentence; we count how often it stops and listens.
Time to stop talking
From the caller's first loud sound to the agent going silent at the caller's ear, median.
lower is better
Exact resume after a false alarm
When a cough or a “hmm” pauses the agent, it should carry on from the exact sample where it stopped.
Card servicing: 80 calls, one run · interruptions: 94 calls after, 59 before · Every figure, by model
Pick the engine per agent
Each agent runs on exactly one voice pipeline. Switch it in the agent's Deploy tab; the prompt, steps and tools stay the same.
- Built in
7Nadi voice engine
Our own engine: speech detection, transcription, the language model and speech on models we host, with barge-in and the safety transition.
1.2–1.7 scaller stops → agent speaks, on a three-step graph test; measured on Gemma 4 E4B, October 2026; 26B results to follow - A second pipeline, on request
O1 engine
A separate engine with its own audio core, connected through the same pipeline interface. It runs prompt and graph agents, tools and extraction. Available on request: we enable it for your workspace.
≈ 1.0 scaller stops → agent speaks, on the same graph test; measured on Gemma 4 E4B, October 2026; 26B results to follow - Your pipeline
Bring your own engine
Register an engine you run. It receives the call's stream over a signed link and reports turns, tool calls and results back.
YoursTwilio, Exotel or our own stream format
Turn times are from phone-less test calls on our machines, three turns each, not a guarantee for your line.
Safe by structure, not by hope
Calls about money, identity and danger need rules the model cannot talk its way around. These are built into the platform.
Verification before any account action
Account steps and their tools sit behind the identity check, so a caller who is not verified cannot reach them. 0 bypasses in our 80 test calls.
Keypad and stepsA safety transition from every step
Danger, an emergency or suspected fraud is checked on every caller turn, before the step's own exits, and it wins.
Safety transitionsAI disclosure
Turn it on and, on outbound calls, the agent says it is an AI assistant right after its greeting. Default lines exist in Hindi and English.
AI disclosureCalling hours, do-not-call and consent
Checked when a call is created and again just before it dials, together with daily caps.
Calling rulesPersonal data masked
On our engine, card, Aadhaar, PAN and account numbers are masked before text reaches the model, the logs or your webhook. Keypad digits are never spoken or recorded.
PII redactionTools, not legal advice
These controls help you follow the rules that apply to your calls; knowing which rules apply stays with you.
What it does not cover
Templates written for Indian businesses
Each one is a full plan: the steps, the fields it fills in, the tools it calls and the rules it keeps.
Credit-card payment reminder
Reminds card holders of the amount due, confirms identity on the keypad and records a promise-to-pay date.
Open the templatePolicy renewal reminder
Reminds policyholders of the renewal date and premium, and records whether they will renew.
Open the templateCash-on-delivery order confirmation
Confirms COD orders and the delivery address before dispatch, to cut fake orders and returns.
Open the templateAppointment booking & reminders
Books, reschedules and cancels doctor appointments on inbound calls and confirms upcoming ones outbound.
Open the templateCandidate pre-screening
Calls applicants, asks the same screening questions of everyone and books the next interview round.
Open the templateKYC / document follow-up
Reminds customers which documents are pending, explains how to submit them and books a callback.
Open the templateThe parts a production call needs
Each of these is in the product today, and its limits are written down in the documentation.
- Our own voice engineSpeech detection, transcription, the language model and speech, with barge-in, on models we host ourselves. Our Hindi voices are built from the open IndicVoices-R corpus (CC BY 4.0).
- Function toolsThe agent calls your HTTP APIs mid-call to look things up or write them back.
- Knowledge basesUpload a PDF or point at a web page; the agent answers from it instead of guessing.
- Batches and campaignsUpload a contact list, or run a workflow that calls, waits and retries per contact.
- HandoverPass the caller to a person's phone number when the call needs one.
- Results by webhookTranscript, summary, extracted fields and cost for every call, pushed to your URL.
- Compliance toolkitCalling hours, do-not-call list, consent, daily caps, AI disclosure and PII masking, checked before dialling.
- Prepaid walletTop up, pay per second of call, and read every charge in a ledger.
What we shipped
In the order it went in. Each entry is in the product today.
Graph agents with a safety transition
Steps with their own exits, and safety transitions offered from every step on every caller turn.
Workflow tests
Scripted callers per graph agent, repeats, flaky flags, coverage and a safety probe.
Voice pipelines, including your own engine
Each agent picks one pipeline: our engine, a second one, or an engine you register.
Agent builder from documents
Attach SOPs, get a template suggestion, edit the draft and accept suggestions round by round.
Simulated calls with recordings
Up to five whole calls through the voice pipeline, with checks, latency and audio to download.
Barge-in, second version
Pause, then decide: real words stop the agent, a “hmm” or a cough lets it carry on from the exact sample.
A new dashboard home
Alerts, the last seven days and your agents on one page, grouped as Build, Test and Run.
Bring your own phone line
Connect your account with a provider and use its numbers, inbound and outbound. While you build, numbers and dialling can be simulated.
- 01
Exotel
Used on real calls.
- 02
Twilio
Tested against simulators, not yet on a real call.
- 03
Plivo
Tested against simulators, not yet on a real call.
Hear it
Get a call from one of our agents
Leave your number. We will call you back.
Start with the call you make most often
Start from a template, test it in the browser, and place a real call when you are sure.