Built for the hardest audio in the world.

01Multi-stage reasoning, not a single-turn bot

A single-stage voice bot maps an utterance to an intent and an intent to a reply. That is enough to ask somebody to confirm an order and press 1. It is not enough for either of the two windows we work, because in both of them the useful information is the thing the buyer has not said yet.

Our agent runs the call as a sequence of stages rather than a lookup. It holds the shipment’s state, decides what still has to be established, asks for that specifically, checks the answer against something outside the conversation, and only then commits an action to your courier.

1. 01TriggerOrder created, or a reason code lands
2. 02PlanWhat this call must establish
3. 03ConverseAsk, listen, handle interruption
4. 04ValidateCheck the answer against the pin code and the order
5. 05DecideApply your rules to the outcome
6. 06Write backCourier and store, inside the SLA

The validate stage is the one buyers underestimate. An address the buyer says confidently and a pin code that does not contain it is the most common way a reschedule fails twice, and catching that costs one lookup before the instruction is committed.

02Latency, turn-taking and interruption

A buyer hangs up on a robot, and what makes something sound like a robot is almost never the voice. It is the gap. A reply that lands half a second late reads as a machine thinking, and the caller starts talking over it, and the call degrades from there.

So the agent targets sub-second response at the point the buyer experiences it, handles barge-in rather than talking through it, and mirrors the pauses of natural speech instead of firing the moment silence is detected. On an Indian mobile call, silence is not the same thing as a finished sentence.

The number we publish is measured from the audio, which is why it is higher than the numbers quoted from server logs. A log can only say when our server started talking. It cannot include the network, the codec, or the carrier. How that measurement works.

03The Indic speech problem

Global speech models run 20–30% word error on Hindi and English code-switching. That is not a rounding problem. On a noisy tier-3 line, “teen baje” becomes “13” and the reschedule is booked for a time that does not exist.

आपका order kal deliver नहीं हो paaya

That sentence is the problem in one line: the frame is Hindi, the nouns are English, and neither a Hindi model nor an English model is reading the whole of it. Code-switching is not an edge case in Indian commerce, it is the default register, and an agent that treats it as an exception fails on the majority of its calls.

Our models are tuned on telephony-grade audio rather than studio recordings, because that is what an 8kHz carrier hands you, and the call is happening next to traffic rather than in a booth.

Languages the agent handles: Hinglish, plus Hindi, English, Bengali, Marathi, Tamil, Telugu, Kannada, Malayalam, Gujarati, Punjabi, Odia.

We are confirming which of these are production ready today rather than listing all of them and hoping. Ask us for the current list and we will tell you which is which.

04Telephony-grade audio

Everything upstream of the model is a lossy Indian mobile call: 8kHz, a lossy codec, packet loss, a second party on speakerphone, and a background that is genuinely loud. A pipeline tuned on clean 16kHz audio loses accuracy exactly where your volume is.

The agent is built to tolerate that rather than to require better input. It also refuses to guess: a call whose audio was too poor to work with is reported as unmeasurable rather than being scored as if it went fine, which is the same rule our measurement instrument applies to itself.

05The write-back layer

This is the part that separates the product from a dashboard, and it is the reason the two windows are worth anything at all. A corrected address that dies inside a report is worth nothing. Without write-back into the courier system inside the SLA, the parcel RTOs anyway and the call was a cost.

Orders arrive from Shopify and WooCommerce. Outcomes go back out to Shiprocket, Delhivery, Ekart, Xpressbees: a booked reschedule, a corrected address, an alternate number, a forced re-attempt on a disputed delivery, a prepaid conversion, or a clean cancellation.

Every write-back is recorded with the time it landed against the time the SLA window closed, because “we called them” and “the courier acted on it” are different claims and only the second one is worth paying for.

06Human handoff

You write the list of situations the agent must not attempt: value thresholds, specific reason codes, an angry caller, anything you would rather a person handled. When one fires, the call transfers to your team live, on the same call, rather than becoming a ticket somebody reads tomorrow.

The agent also stops rather than improvising when it does not have the information a decision needs. Where the V2 site would have said the agent decides, this one says your rule decides and we execute it, which is both truer and the thing an operations lead actually wants.

07Security and data handling

We process order data, phone numbers, call audio and transcripts. Recordings and transcripts are retained so that every call can be measured from its own audio and so that you can listen to any call we bill you for.

Retention periods, deletion requests and the position on TRAI, DLT registration and permitted calling hours are set out on the privacy page and the calling policy page, including the parts still under legal review. We would rather point at an unfinished review than imply a finished one.

08The specification, in one list

• Sub-second response latency
• Barge-in and interruption handling
• 11 Indian languages + Hinglish
• Telephony-grade audio, 8kHz tolerant
• Multi-stage reasoning, not single-turn
• Automatic re-attempt scheduling
• Live human handoff
• Call recordings and transcripts retained

We are checking each line above against a live call before we let it stand as a specification. Ask us to demonstrate any of them on the call we book with you.