SquadStack

Voice AI Agents that drive revenue by giving every lead the best pitch, on their preferred channel, in their language.

[SquadStack @SquadStackAI]

India's largest enterprises are sitting on a ₹50,000 crore revenue leak.

Nobody in the boardroom is talking about it.

Here's what's actually happening and why fixing it is the biggest revenue unlock left on the table:

[Apurv Agrawal @apurv_agrawal]

When a customer says “chhe hazaar nau sau” on a call, they mean ₹6,900.
One global model heard ₹1,900. Another heard ₹900.

In a sales conversation, a mistake like this can change the outcome entirely.

This is one of the reasons we built Arth v2, our in-house streaming speech-to-text model.

[Apurv Agrawal @apurv_agrawal]

A misheard word on a sales call can mean a missed conversion, a wrong payment date, or a lost sale.
For us this isn't just a technical problem but a business problem.
This is why we built Arth V2 from scratch built for the realities of India's sales calls. https://t.co/wdeNwZLfdB

[SquadStack @SquadStackAI]

Then it finds the exact change that would have turned more of those calls into outcomes. One change at a time. Each one tested on live traffic before it scales.

The agent you launch with is the floor. The calls it makes are what move it up from there.

[SquadStack @SquadStackAI]

𝘔𝘶𝘫𝘩𝘬𝘰 𝘣𝘩𝘪 𝘵𝘰 𝘭𝘪𝘧𝘵 𝘬𝘢𝘳𝘢 𝘥𝘦, 𝘵𝘩𝘰𝘥𝘪 𝘴𝘪 𝘵𝘰 𝘭𝘪𝘧𝘵 𝘬𝘢𝘳𝘢 𝘥𝘦.
- 𝘌𝘷𝘦𝘳𝘺 𝘈𝘐 𝘢𝘨𝘦𝘯𝘵 𝘴𝘵𝘶𝘤𝘬 𝘢𝘵 𝘵𝘩𝘦 𝘴𝘢𝘮𝘦 𝘤𝘰𝘯𝘷𝘦𝘳𝘴𝘪𝘰𝘯 𝘳𝘢𝘵𝘦 𝘧𝘰𝘳 3 𝘸𝘦𝘦𝘬𝘴

[SquadStack @SquadStackAI]

It reads the calls your team never gets to. Finds the near-misses where the customer almost booked. Finds the drop-offs where something broke in the first 30 seconds. Scores each one on whether it converted and how cleanly the agent handled it.

[Vikas Gulati @_vikasgulati]

Arth v2 is live.

It's our streaming speech recognition model, trained from scratch on 600M+ minutes of Indian sales calls.

On 863 real Hinglish sales calls (8 kHz phone audio):
- 67 ms to final transcript, the fastest engine we tested
- 8.85% semantic WER, matching the best streaming engine
- Error rate rises just 3.8 points from clean to noisy calls, the smallest climb of any engine we tested

Every speech model looks good on a quiet line. Our customers pick up at a shop counter or in a moving auto, switching between their regional language and English mid-sentence. The team ran 100+ ablations across pre-training and post-training to get here.

The benchmark is public, call by call, with no overlap with Arth's training data: https://t.co/nWmOrFNc85

Arth is the first model from our model factory. Lift used to improve only the harness. Now it improves the weights too. Calls where Arth misheard a word become training data for the next version, which ships every few weeks.

More models are coming soon.

https://t.co/EOcqI5QcW9

[Apurv Agrawal @apurv_agrawal]

Introducing Arth V2, SquadStack's in-house streaming speech model, built for India's phone calls 🧵

At 67 ms, it's the fastest model that outpaces every engine and beats Sarvam, Deepgram, Google and OpenAI on accuracy across 863 real Indian sales calls.

[Apurv Agrawal @apurv_agrawal]

Our AI agent did its homework before this seller even picked up.

It knew what he sells, where his shop is, how many buyers want products like his and where they are, what a typical order looks like, and how many enquiries similar sellers nearby are getting.

[Vikas Gulati @_vikasgulati]

Lift update: personalisation.

Lift already seeds the best baseline agent from your knowledge base, recordings and our knowledge pool, then keeps improving it from live calls.

Now it also writes a brief on every lead for the agent before the dial. Result: 1.5x more conversions with personalisation alone. 📈

[Apurv Agrawal @apurv_agrawal]

Voice AI pricing in India today looks like BPO pricing. Vendors compete on cost per minute. Buyers compare rate cards. Everyone is negotiating an input.

No lender walks into a board meeting and says “we bought 4 lakh minutes of voice AI this quarter”. They bought disbursals. Insurers bought renewals. Universities bought enrolments.

Cutting calling costs by 60% saves about 2% of revenue. Lifting conversion by 10-15% adds 10-15% of revenue. The industry has been pricing on the wrong number.

Today we announced outcome-based pricing for our Voice AI sales agents. A low fixed base to run the agents, and a success fee on the revenue they create.

We don't sell voice AI. We own a line in our customer's MIS.

https://t.co/UMyVUUNc4O

[Vikas Gulati @_vikasgulati]

In July we launched Lift: self improvement loop for our Voice Agents.

Today @SquadStackAI we are announcing outcome-based pricing. A low fixed base, and a success fee on the revenue our agents produce. If the agent doesn't convert, we don't earn.

Per-minute pricing pays voice AI vendors whether or not the call converted. It turns voice AI into a BPO with better margins.

The right metric is cost per outcome. Lenders should pay on disbursals. Insurers on renewals. Universities on enrolments.

https://t.co/WnppPiSd9R shifts voice AI pricing from minutes to sales outcomes https://t.co/3Wk7DvNksg via @economictimes

[SquadStack @SquadStackAI]

Today we announced outcome-based pricing for our Voice AI sales agents. A low fixed base to keep the agents running, and a success fee tied to the revenue they produce.

Live now for banking and financial services, more sectors coming soon.

https://t.co/lc9C4HOqPK

[SquadStack @SquadStackAI]

Lenders buy disbursals. Insurers buy renewals. Universities buy enrolments.

Enterprises don’t buy minutes. Yet that is how voice AI is priced in India.

[Apurv Agrawal @apurv_agrawal]

https://t.co/nAr09UHMKX

[Apurv Agrawal @apurv_agrawal]

https://t.co/kuaFb17LOR

[Apurv Agrawal @apurv_agrawal]

Day 5 of 5: Introducing Inline PII Redaction

@SquadStackAI redacts customer PII in real time, before it reaches the LLM.
The calls that drive the most revenue in BFSI are the ones where customers speak Aadhaar numbers, PAN, bank details out loud.

In most voice AI systems, that data sits in plain text across the full stack. STT output, transcript, LLM prompt, model provider's API, logs, QA dashboard. No redaction at any step.

The hard part: PII on voice calls doesn't arrive in one clean line. A customer spells their PAN across four turns. They switch between Hindi and English mid-sentence. Single-turn detection misses most of it.

So we built inline PII redaction, a 6-turn rolling window between the STT engine and the LLM that catches PII split across fragments. 16 entity types, tuned for Indian voice conversations.

A regulated brokerage's compliance team recently reviewed this and signed off for production at scale.

To learn more: https://t.co/q5TcJjYOnB

[Apurv Agrawal @apurv_agrawal]

Day 4 of 5: Introducing RAG & Context Management

Your AI agent can now handle long complex calls without a human stepping in!

A few weeks ago I wrote about how we fixed 20-minute long revenue calls. How the agent could remember everything the lead said and hold up!

Remembering wasn't enough on its own. In some instances, the agent just couldn't reliably answer the very hard and specific product question the entire call was building toward.

Not because the knowledge was missing. Every policy detail, pricing sheet, FAQ was loaded into the prompt. The prompt was just so heavy that the model skimmed past the exact answer the caller needed.

So we moved product knowledge out of the prompt entirely, and put it in a knowledge base.

The agent reads from it, only when the caller asks. Retrieval takes under 100 milliseconds. On blind tests, including an 87-page product manual, it finds the right answer 97% of the time.

Context Management gave the agent memory. The knowledge base gives it depth.

Together, the 20-minute insurance call, the loan advisory call with fifteen branching questions, now close without a hiccup or the need of transferring to a human.

Full story: https://t.co/vKc5maMdau

[Apurv Agrawal @apurv_agrawal]

Day 3 of 5: Introducing Overwatch

A question I hear from leaders running voice AI at scale: how do we know what's actually happening on the calls?

The most expensive problems never throw an error. If the greeting has a two-second lag, or the agent fumbles a rebuttal on a specific objection, or it marks a lead as qualified without capturing the right entity. You find out days later, when the numbers slip.

That's why we built Overwatch.

A call slows down? Overwatch tells us whether it was the STT, TTS, LLM, or a tool call.

A campaign stops connecting? We see the telephony vendor, the drop reason and the DND status before anyone files a ticket.

A metric spikes? You go straight from the number to the exact calls it touched, by ID, ready to open and listen to.

Every problem you miss on the first call repeats on the next thousand. At ten thousand calls a day, seeing it early is the difference between scaling and quietly leaking revenue.

Overwatch is now live inside our Humanoid Voice AI Agent.

Full story here: https://t.co/KxVz1UjK3H

[Apurv Agrawal @apurv_agrawal]

Day 2 of 5: Introducing Vernacular Voice AI Agents

Every voice AI platform lists a dozen Indian languages on the feature page. The count keeps climbing.

But deploy these agents in production and the conversations fail. The agent speaks “correct” Tamil. It sounds like a textbook. The customer knows in three seconds. They hang up.

We found that naturalness splits roughly 20/80 between TTS configuration and dialogue engineering. The 80% is the part nobody wants to do.

We mine fillers from real human agent calls, test every filler against every voice on real 8kHz telephonic calls, maintain a kill list of phrases that signal AI, and write every line of dialogue by hand.

No LLM generates this. That's the point.

We've launched vernacular voice AI agents across India's top languages.
They’re live across large-scale enterprise campaigns.

We wrote about how we built it: https://t.co/f9mGRaMUX2