Shahriar Tajbakhsh, Co-Founder of Metaview, on how AI is solving hiring’s data problem

Carlos: What did you believe in 2018 that turned out to be right, and what did you get completely wrong that shaped how you were building Metaview at the time?

Shahriar: I think the fundamental secret was that Sahil and I spent a lot of time at top-tier companies building teams — not as recruiters, but as operators, as engineers and product people. And even at those high-caliber companies, it was clear how easy it was to get frustrated by the lack of precision in the process. So we started thinking about how we could make it such that no bad hiring decisions are ever made. That was the founding question before we actually founded the company. And the core thesis was: it’s very hard to create that future if you have no data to base your decisions on. Hiring at the time was a series of disjointed conversations by different people with very little information sharing. We thought hiring decisions are the most important decisions companies make, and that conversations constitute ninety percent of the valuable data in this process — but they weren’t being captured or operationalized, and no learnings were being drawn out of them. That’s what happened in 2018. We said if we could record these conversations and transcribe them, we could then do stuff with them. We spent about four years trying to do stuff with those conversations, in what I would describe as fine, but non-game-changing ways. The core thesis never changed. And it was very helpful that we got lucky at the right time, with LLMs being the perfect technology for our kind of data. We stayed alive long enough to get lucky.

Carlos: Which of the cultural things you brought from Palantir and Uber became an unfair advantage at Metaview, and which ones did you consciously leave behind?

Shahriar: I have an almost unhealthy obsession with talent density and the quality of the next person that joins the company. That’s Palantir DNA — this obsessive desire to hire exceptional talent who might be weirdly weak in certain aspects but unreasonably spiky in one thing. That spike presents itself in magical ways.

The other thing is communication. If your talent density isn’t right, you’re dead on arrival. Once you do get exceptional people in, they need to have all the context to make good decisions — otherwise everything’s blocked on a founder, which is hugely inefficient. We have a philosophy internally called PUP, which stands for Public Unless Private — every form of communication within the company is public to anyone in the company. We do not have direct messages on Slack. Every channel is public and available to everyone. Of course, if something’s personal, you take that private, but the default posture is public.

That gets more interesting on email. Every email that’s sent is also available to everyone else, particularly powerful with customers. If I’m a new account executive who joined today, I can go and look at every single piece of communication anyone at Metaview has ever had with a given customer, from 2018 until now. All meetings are recorded and transcribed, and the data is available to everyone. The thing we did not anticipate is that it also results in exponentially good outcomes with LLMs, given that all the context is now available for them to ingest. It’s not that we designed it with that in mind — it’s just a huge upside.

Carlos: Can you give me an exact example of where the line between copilot and autopilot has moved?

Shahriar: Let’s take two worlds — a legacy company and an AI-built company. In a legacy company, when they want to reach out to a candidate, they create a sequence in whatever tool they have. Every candidate gets an email, subject such and such, body such and such, with one customizable line about their background. That roughly results in the crappy emails you get, where it says something like “I noticed you’ve been an investor for fifteen thousand years” — pure boilerplate. Three days later, everyone gets the same follow-up, no matter their background or what they do.

In an autonomous world, which exists today, the agent takes each individual candidate and does deep research on them across the entire internet — everything they’ve worked on, where, what stage the companies were at, what headcount to what headcount, what revenue to what revenue. It listens to their podcasts, reads their papers, reads their blog posts. That results in a multi-page dossier — the most comprehensive document available on that person online. Then the agent decides, given how much it knows, the best way to engage them for that particular role. It determines a subject line and content for the email or the LinkedIn message that’s precisely designed for that person and no one else, reflecting the fact that hours have been spent paying attention to them.

It makes no sense to send a fixed follow-up three days later on a schedule — because while the deep research was running, we might notice this person is clearly online every Sunday, so the best time to reach out is actually via an X DM on Sunday saying, “I dropped you an email, can you check it out?” You can get hyper white-glove with each individual, but this can only happen in an autonomous world, and it’s clearly going to outperform humans, because no human has the time to do any of this. If a human was to pay this much attention to every candidate, they’d need 120 years.

Carlos: If you were starting Metaview today, what would you build first?

Shahriar: It’s clearly not going to be the notetaker. It was very novel and very hard to do three or four years ago; it’s pretty easy to do something good enough today. There’s still a lot of value in making it vertically amazing — a notetaker specifically designed for recruiting will be very good at that, one designed for doctors will be very good at that — but I just don’t think that’s the first product today. I think the first product today would be what we’re building now, which is Fillmore — our autonomous AI coworker that goes from context to first calls booked with the candidate, fully autonomously, with no human intervention. I expect that to be the most impactful product this industry has ever seen. Clearly that’s what we would do if we started today.

Carlos: How did you manage to get skeptical enterprise buyers to stop looking at this as an AI novelty and trust it with mission-critical workflows?

Shahriar: We have a lot of work to do, obviously, always, to build more trust. So far we’ve done well. Part of it is just through unreasonable hospitality — listening, caring, doing things people want us to do, even if our interpretation of what they want is a little different. I think it’s just through caring and obsession. We do plenty of things that don’t work, and we turn a lot of those into things that do work, through listening and observing how people use the product.

Sahil and I have been working on this problem since 2018. I bet it’s a lot of hours of just caring about one thing, one domain, one industry — talking to people, learning about them. That’s how you form intuition about what the problem is and how you want to solve it. That meets the real customer, and four feedback iterations later, you do the thing that resonates. That builds trust — listening and building the thing that actually solves the problem.

But also, ultimately, you either change the game for someone or you don’t. We have a track record of mostly changing the game for folks, and every now and again we do stuff that doesn’t work, and we move on and learn from it. That compounds. The AI notetaker earned us the right to do more. We did that well, kept learning from it, improving it, and that earned us the right to build the next product. When we had our next product, a lot of people were almost blindly willing to try it, because it had our reputation on it.