There is no shortage of startups trying to develop enterprise AI assistants. Although less common, there are AI assistants that can actually perform tasks for multiple work apps at the same time. That's the promise of Narada AI, a startup building an AI assistant based on new research from the University of California, Berkeley.
Narada has been operating in stealth mode for two years, making its public debut on stage today as part of Startup Battlefield 20 at TechCrunch Disrupt 2024.
Two of the co-founders, Kurt Keutzer and Amir Gholami, co-authored a paper earlier this year on the LLM Compiler, an AI system that performs multiple functions simultaneously. Their startup is primarily based on this open source technology, and they believe this is a key differentiator from the general purpose AI chatbots out there.
Dave Park, co-founder and CEO of the startup, says his team used this as the basis for building custom AI models that can be used with productivity tools. Park, who earned a PhD in computer science from Stanford University and spent 24 years in enterprise sales, believes the LLM compiler and Narada's ability to use websites without APIs are the company's “secret to winning the enterprise agent race.” I believe that it is the source of
This idea sounds promising, but how does the agent actually work? In fact, we've seen that the assistant can successfully perform several different tasks using AI generated through various work apps. , ultimately saving seconds or minutes in various parts of the day.
The Assistant appears in a separate chat window in your browser and can draft emails, invite you to your calendar, take notes for meetings, and search the web for you. The company says its assistants can also work with enterprise applications, such as searching for invoices in SAP, taking notes on video calls, and analyzing information from Salesforce's many apps.
I asked the AI assistant to craft a friendly-sounding email to decline an invitation I received. Within seconds, your drafted email appears in Gmail, with the correct recipient (even though you didn't give them their email address, it found the correct one), subject, and body all filled out, with the bottom has my signature written on it. All I had to do was review the content and click send.
At another point, I prompted the AI assistant to find a highly rated Japanese restaurant in my San Francisco neighborhood and book a dinner invitation with a friend on my calendar at a time that worked with my schedule. I found a restaurant, created a calendar invite, and drafted an email to a friend with that information.
So how do agents do all this?
To use email and calendar, agents use APIs in part to access these programs through a developer backend. But Park says AI agents are also clicking, scrolling, and typing on the front end of websites (for example, this is how you open an email draft in Gmail). This front-end agent, called Web Redemption, allows Narada to use enterprise applications like HubSpot without an API.
Gholami, the company's CTO, said the agent works like a Roomba, creating an internal map to understand new websites and applications. When a user tells Narada that they want to use a new application, the agent plans it and helps them understand how to use it. That's the idea the founders presented to me.
This is how Narada looks in your browser Image credit: Narada AI
But Narada isn't the only startup trying to create AI agents that can use websites through the front end. This is similar to the idea behind Anthropic's computer usage and Rabbit's LAM. However, these agents are difficult to implement in practice and require significant maintenance to keep running. If the web page layout is updated, the agent may stop working.
The main difference with Narada's agent is that it is focused only on enterprise applications, rather than being a general-purpose agent for any website. (When I tried to use Narada for LinkedIn or Facebook, I received an error message. However, the company's website has a demo that allows engineers to use the tool on LinkedIn.)
When it comes to LLM compilers, it appears that industry players have already implemented an open source approach. Gholami told TechCrunch that LangChain and LlamaIndex are already integrated with the LLM compiler. However, Narada's tools differ from these tools in that they are enterprise-focused. The startup already uses agents for Fortune 500 companies, but won't say which ones.
So, is this a replacement for a real-life assistant? Not much. However, the tool sometimes felt like it was using shortcuts for mundane tasks. This cannot be said about many AI tools today.
What made me a little uncomfortable was how much access I had to give this AI assistant. Narada can read all my emails, see my entire calendar, and knows my entire contact list.
As with any such “smart assistant” or helper app, you need to trust not only the technology but also the company itself. You must trust that Narada will not misuse your data or your company's data. That said, the company has promised not to train its AI models based on customer data.
Narada says it has raised millions of dollars so far from several advisors it has brought in, but the CEO is now considering raising more money from traditional venture capital. It says that there are.
Narada AI demo.
Source link