@DecagonAI is building human-level AI support for customer-facing teams.
@natarajsindam sat down with @AshwinSreenivas on to talk about
1. scaling to $1.5B in 2 years
2. approach to finding PMF
3. evolution of voice agents
& a lot more.
Snippet here, full links in comments
Record yourself speaking.
Watch it back.
Cringe.
Repeat.
Varun Puri says that's the single best way to improve communication — no shortcuts.
Also: speak in threes, start with a hook, use alliteration.
Full convo in the links on profile.
Varun Puri worked directly for Sergey Brin and realized something: even the most powerful people are just guessing.
"The world is run by blind people leading the blind."
Saying "I don't know" isn't weakness. It's honest.
Full convo in profile links.
Building is easier than ever—so proving value is harder than ever.
No moat from connections or access anymore. Just one question:
Do customers use it? Every day?
Varun Puri lays it out plainly. Links in profile.
"LinkedIn is cringe—but it's our #1 lead source."
Yoodli's Varun Puri admits he used to mock people who posted constantly. Now he's one of them.
His take on why founder-led content beats company pages every time.
Full convo in profile links.
Textbooks → AI that talks back, corrects you, and certifies you.
Varun Puri calls it "experiential learning" — and thinks it goes way beyond sales training.
Fascinating direction for how we actually learn things.
Full convo in links on profile.
AI roleplays are getting commoditized—Varun Puri knows it and doesn't care.
The real bet is the full experiential learning stack: deep customization, enterprise integrations, compliance, UX.
Prototypes are easy. Depth isn't.
Full convo in profile links.
AI handles 90% of execution.
The real edge? Judgment.
Knowing what's good vs. bad—that's the skill that actually matters post-AI.
Varun Puri breaks it down.
Full convo in the links on the profile.
Google asked Yoodli for a feature that didn't exist yet.
Varun Puri said yes anyway, then built it.
That one call turned into their entire enterprise product.
Full conversation in the links on profile.
When AI gives everyone the right answer instantly, what's left?
Varun Puri says it's how you communicate—your authenticity, vulnerability, relatability.
The most human skill becomes the rarest edge.
Full conversation in profile links.
Building an AI data center in the middle of nowhere?
You might literally have to build a town around it.
Stranded power, labor shortages, local protests — the bottlenecks are wild.
Full convo in the links on profile.
A Databricks/MosaicML demo showing CUDA code running straight on AMD MI250 was the lightbulb moment for TensorWave.
That one proof sparked the entire Beyond CUDA movement.
Full conversation in the links on profile.
"If it works, we'll encourage them to use AMD."
That one line was enough. TensorWave was formed the next day.
Full story in the links on the profile.
Why should AI companies be stuck with whatever Jensen decides to sell them?
Jeff Tatarchuk wants TensorWave to be the "AI utility company" — reliable AMD compute you don't have to babysit.
Full conversation in the links on profile.
$100M raised. Still eating ramen.
Jeff Tatarchuk explains why AI infrastructure eats capital alive—cost of capital, credit ratings, power timing.
The math is brutal.
Full convo in the links on profile.
What does success look like for AI infra?
Jeff Tatarchuk's answer: when customers stop thinking about it entirely.
Full conversation in the links on the profile.
"You only need CUDA if you don't understand low-level GPU development."
TensorWave's Jeff Tatarchuk on NVIDIA's margins, 150 neo clouds that can't all survive, and why the power grid is the real AI bottleneck.
Full convo in links.
150 neoclouds trying to survive when there aren't even 150 AI labs.
Jeff Tatarchuk breaks down why "plug in GPUs and rent to OpenAI" isn't a business model.
Full convo in the links on profile.
Gong's CPO Eilon Reshef: "Ignore the AI hype. AI won't replace PMs, but it should upgrade how you work."
Use it as a thought partner for prototypes, docs, ideation. Real challenge? Nobody's cracked where AI truly fits in software.
Full convo in bio.
Speech-to-text in 2015: 30% word error rate
Gong literally hid their transcripts behind 4 clicks because they were unreadable
They had to build their own engine just to make it work
Now we take accurate transcription for granted
Full convo in profile