Lindy is building your first AI employee 💻
Plus: Flo on AI agents, Lindy's team and the n8n integration...
Published 15 Feb 2024CV Deep Dive
Today, we’re talking to Flo Crivello, founder of Lindy.
Lindy is a no-code tool for building your own AI agents or ‘AI employees’ - helping you automate tasks like email drafting, scheduling meetings, note-taking, customer support and more. It's similar in UI to ChatGPT, but includes 3,000+ new integrations with n8n.io and is activated by external event triggers - such as a new email received by your email client.
Today, Lindy is opening up after months in private beta. It has become a go-to AI tool for founders, entrepreneurs and small business owners, and the startup is backed by Menlo Ventures, Battery Ventures, Coatue, Jack Altman, Elad Gil and more.
In this conversation, Flo walks us through his vision for Lindy, building AI agents, and his goals for 2024.
Let’s dive in ⚡️
Read time: 8 mins
Our Chat with Flo 💬
Flo - welcome to Cerebral Valley. First off, give us a bit of background on yourself and what led you to start Lindy?
I'm a software engineer by training. I was a software engineer for ten years, and I still consider myself one — building companies is just designing one big system. Previously, I worked at Uber for 5 years as a product leader in various divisions.
After Uber, my first startup was called Teamflow, building a virtual office for remote teams. When GPT-3 came out, we got super excited about the new technology and were like “hey, what can we build with this?” First, we built a meeting recording bot that would record and summarize meetings in the virtual office.
We also had salespeople at the time, and they were like, “I hate having to update Salesforce after every sales call, can it do that for us?” And we were like “yeah, no problem!" So we did that - and then our engineering team was like “could it also update the Linear task board after our sprint planning every Monday?” And we're like “yeah, sure!”
So, as we built out this system, we realized we could get these language models to write the API calls for us. This was mid-2022 before anyone was really talking about agents - before LangChain, even. Little by little, we realized this was the opportunity - building an agent that can take action for you in the world.
How would you describe Lindy today to someone who has never used it? And who is finding the most value in it so far?
Lindy’s a no-code tool for you to build your own AI agents or AI employees. It's like ChatGPT, except it has 3,000 integrations out of the box instead of requiring you fiddle with OpenAPI specs, and you can set it up with triggers so it’s woken up by external events, like a new email arriving via Gmail.
Today, we’re mostly used by founders, entrepreneurs and small business owners.
Which insights led you to pursue building AI agents, before that concept became popular?
It was really building that meeting recorder that, little by little, had to integrate with systems like Salesforce, HubSpot, Gmail, Linear etc. At first, we built these integrations manually, but over time, we started to see a pattern emerge and realized they’re all just API calls.
We built a modular architecture where you build all of these different tools and feed them in the context window of these models — which side note, at the time was only 2,000 tokens, it still blows my mind we even tried to build this when those were the model capabilities!
That's when we really realized these large language models can do a lot more than just generate copy for you. They can actually take action for you.
And, for all the talk about “generative AI,” I got a lot more excited about “agentic AI” — AI that does stuff for you. Our GDP isn’t made of copywriters or illustrators - it’s made of work and actions.
Walk us through Lindy’s major use-cases today.
Today, creating custom chatbots is definitely one of the bigger use-cases. We also have a couple of use-cases that have really surprised us - one of them being people using Lindy as a life coach. Lindy is helping them keep their lives on track, hit their goals and improve themselves, which I think is pretty awesome.
Another use-case that we really didn't expect is from the health world — today, lots of doctors use Lindy as a medical scribe. We transcribe consultations with Whisper, and Lindy generates notes in a special format used by doctors.
Lastly, lots of founders are using Lindy to automate their customer support, plugging her to their Gmail, Intercom, Zendesk and more.
But really, we’re constantly surprised by what people build with Lindy — you can really build anything.
How does Lindy personalize to each user? And how does it handle specific errors or edge cases that come with that?
Lindy learns about you the more you interact with her. The parallel would be: in ChatGPT, you have thumbs-up and thumbs-down buttons at the bottom, but they don’t actually do very much. Lindy has the same feature, but if you thumbs-up a query, we embed it in a vector database unique to your account. For future queries, we retrieve the closest-such examples that you've previously thumbed-up, and inject them into the context window.
A friend calls it the ‘poor man's RLHF’, but it works surprisingly well. If you thumbs-down, Lindy asks you “what happened?” You give feedback, which Lindy uses when she tries again — and you can do that again and again. Once Lindy gets it right, which at some point she does you give a thumbs-up. And then the next time you ask something like this, she just gets it right. That’s a super easy way for you to get Lindy to do what you want to do.
In terms of errors — the simplest way to rectify these is by changing the Lindy's guidelines via her settings. We do offer a safe mode that's turned on by default, which means that Lindy can only perform actions with side effects — like sending an email — by asking for your permission first. This is so she shows you what she's about to do before she does it. While you're still building up confidence in the system, we suggest you leave safe mode turned on, and only turn it off once you've iterated enough on your Lindys.
What are some of the major technical challenges with building in the space of AI agents, and how do you balance productization vs. research?
The way we think of our role in the industry is: we’re focused on the cognitive architecture, and on the whole product. We don't build the LLM - we build the layers on top which lets the LLM be useful. These layers let you interface, ingest perceptions and put out actions in the world — and that's the layer that we're focused on.
That layer of cognitive architecture is also a very active field of research right now, and we’re in a pre-paradigmitic world about it — no one knows the right cogarch (cognitive architecture) to take. So, that's really the technical difficulty that we have to deal with. We don't give too many details about ours - that’s a lot of the secret sauce - but we frequently see a paper that everyone's impressed with, which we've been implementing for the past six or nine months.
Generally I think people underestimate how far academia has fallen behind actual SOTA [state of the art]. I was chatting with a friend at OpenAI who was telling me the same thing — it's almost dispiriting that the current state of research is that it’s no longer the papers on Arxiv or in academia that are the frontier. You can only find the frontier in Signal groups with AI folks now, or by grabbing a drink with them — and these don’t publish nearly as much as they used to, if at all.
You just announced an integration with n8n.io - a well-known workflow automation tool. Tell us about why this is a major step forward for Lindy?
Yes! So n8n is like an open-source Zapier. We just announced a partnership with them, in which all n8n integrations are going to be natively built into Lindy. In total, this gives us 3,000+ actions inside Lindy, which is incredible. As far as I know, Lindy is going to be the only AI agent platform that has this many actions. The other ones — even GPTs — ask you to fiddle with open API specs, but in Lindy, you can just add an action and select whichever tool you want to use.
Lindy also offers triggers. To my knowledge, again, we’re also the only ones supporting an external event that wakes up your Lindy. So, if you want to use Lindy for customer support, you could create a trigger every time you receive a new support ticket on Intercom.
And the thing I’m most excited about is that Lindys can talk to each other. It’s almost like OOP [object oriented programming] for agents, where you can get your Lindys to talk to one another and create a Lindy that's basically a function that another Lindy uses. And, because you can publish your Lindys to the Lindy store, you can download another Lindy that someone used and integrate it with a Lindy that you created, and so on and so forth.
How do you think about the desired evolution of Lindy over the next 6-12 months? And any plans to introduce monetization?
Firstly, we’re soon releasing iOS and Android apps that I'm really excited about. I’m also excited about template monetization — letting creators publish Lindys for others to use. And finally, I’m very excited about the work we’re doing on web browsing — actually letting Lindy manipulate and take action within a web browser.
More broadly, I think people underestimate the extent to which our industry is driving towards AGI and how close we’re getting. In 2 or 3 years, we’ll have a reasonably general knowledge worker that can perform any basic task.
In terms of monetization, we’re actually looking at a very different approach from the revenue-sharing model that OpenAI is taking with GPTs. You can almost think of it more like Substack versus Medium. Medium uses a revenue-sharing model, where you put your article in a black box, and then money somehow comes out of a black box. On Substack, however, you can sell a subscription to your own audience or customers.
We're taking the Substack approach, where you’ll be able to decide “hey, this Lindy costs $5 to install or $5 per month”, and then charge your users directly.
What would you say are the Lindy team’s core strengths? And what do you look for in prospective hires?
Not to pat myself too hard on the back, but I think we've done a pretty good job of creating a culture of excellence and intensity and deep care for creating something awesome and very deeply impactful. You can see it in our cultural values - one of them is “you will do the best work of your life at Lindy”, just because we care so much more.
Another value is “we put extraordinary effort to build an extraordinary company”, which basically means we work super hard. We recently lost someone who just wasn't down to work that hard — I’m never glad to see people go, but I’d much rather we only have people who are aligned with our culture. Even when we hire, I actually try to discourage people a bit — I'm like, “are you sure? This is going to be hard.” And I really want to give people an idea of that.
You’ve had some pretty vocal views on remote work. Tell us where you stand with that in terms of Lindy.
So my previous startup was focussed on building software to help remote teams work together. It's not even that we ‘tried’ — getting remote to work was literally the mission of the company. And it’s crazy, but we couldn't figure it out. I'd love to take the blame for this, but frankly, we had a lot of competition in that field — dozens of competitors, and not a single one of them figured it out — which I do think is very telling. Anytime you say that, you've got the remote mob coming after you — but I think it’s so telling that they always mention the two same companies, right? GitLab and Zapier. .
Also, people used to mention Invision, which I actually is a perfect A/B test. There was Invision, and there was Figma — which was very in-person. I think I heard that employee #11 or so at Figma wanted to move to LA to spend more time with his girlfriend, and they let him go. It’s just a fact that people need to see people — it's much easier to collaborate together, to whiteboard and to build a strong culture.
How has your early experience at Uber impacted the way you lead at Lindy today?
Uber was really good at taking 25 year olds and giving them insane responsibilities and autonomy. When I arrived at Uber, I was 23 — and they gave me huge projects worth millions of dollars! I was like, “me?!”. They would encourage me to take major decisions, and that's just how Uber rolled.
That’s the main thing I learnt there — you can take young, hungry, driven talent, give them huge responsibilities, throw them into the deep end, and oftentimes they’ll do wonders. Even if they’ve never done it before.
This is what I like most about this job: finding extremely talented people and giving them an absurd amount of ownership. At Uber, I really learned that if you have very high expectations for people and give them that ownership, they may make mistakes, but by and large, they’ll achieve amazing things. That's really how Uber grew so fast, and so that's also how I run Lindy.
Conclusion
That’s a wrap for our Deep Dive with Flo! Follow him on X to keep up with his work at Lindy.
Read our past few Deep Dives below:
If you would like us to ‘Deep Dive’ a founder, team or product launch, please send us an email at newsletter@cerebralvalley.ai or DM us on Twitter or LinkedIn.