Rapid Engineering Startup Vellum.ai Raises $5 Million as Demand for Generative AI Services Soars

Image credits: Ole_CNX/Getty Images

This morning, Vellum.ai said it closed a $5 million seed round. The company declined to share who its lead investor was for the round, other than noting that it was a staged firm, but told TechCrunch that Rebel Fund, Eastlink Capital, Pioneer Fund, Y Combinator and several angels they took part in the round. .

The startup first caught the eye of TechCrunch during Y Combinator’s latest demo day (Winter 2023) due to its focus on helping companies improve their generative AI suggestion. Given the number of generative AI models, how fast they’re progressing, and how many business categories appear poised to take advantage of large language models (LLMs), we liked his focus.

According to metrics Vellum shared with TechCrunch, the market also likes what the startup is building. According to Akash Sharma, CEO and co-founder of Vellums, the startup now has 40 paying customers, increasing revenue by around 25-30% per month.

For a company born in January of this year, that’s impressive.

Normally in a short update on financing of this type, I spend some time detailing the company and its product, focus on growth, and move on. However, since we’re discussing something a bit nascent, let’s take our time to talk about rapid engineering more generally.

Building parchment

Sharma told me that he and his co-founders (Noa Flaherty and Sidd Seethepalli) were employees of Dover, another 2019-era Y Combinator company, working with GPT 3 in early 2020 when its release was released. beta version.

While at Dover, they built Generative AI applications for writing recruiting emails, job descriptions, and the like, but noticed they were spending too much time on their prompts and couldn’t version the prompts into production, or measure them. the quality. They therefore needed to create tools for tuning and also for semantic research. The huge amount of manual labor was adding up, Sharma said.

This meant that the team was spending time designing internal tools instead of building for the end user. Thanks to that experience and the machine learning operations background of its two co-founders, when ChatGPT was released last year, they realized that the market demand for tools to improve the suggestion of generative AI would be grown exponentially. Then Vellum.

LLM workflows within Vellum. Image credits: Parchment

Seeing a market open up new opportunities to build tools is nothing new, but modern LLMs could not only change the AI ​​market itself, they could make it bigger as well. Sharma told me that until recently released LLMs it was never possible to use natural language [prompts] to get results from an AI model. The switch to accepting natural language input makes the [AI] much bigger market because you can have a product manager or a software engineer [] literally anyone who is a ready engineer.

More power in more hands means more tools are required. On this topic, Vellum gives AI prompters a way to compare model output side-by-side, the ability to search for company-specific data to add context to particular prompts, and other tools like testing and version control that companies might like to ensure that their suggestions are spitting things correct.

But how difficult can it be to apply for an LLM? Sharma said: It’s simple to create a prototype based on LLM and launch it, but when companies end up taking something like [that] to production, they realize that there are a lot of edge cases that crop up, which tend to give weird results. In short, if companies want their LLMs to be consistently valid, they will need to do more work than just skinning the GPT outputs from user queries.

However, this is a bit general. How do companies use refined prompts in applications that require timely engineering to ensure their outputs are well optimized?

To explain, Sharma pointed to a ticket support software company that targets hotels. This company wanted to create a kind of LLM agent who could answer questions like: Can you make a reservation for me?

It first needed a prompt that would function as an escalation classifier to decide whether the question should be resolved by a person or by the LLM. If the LLM was going to answer the question, the model would need to extend the example here by just being able to do it correctly without hallucinating or going off the rails.

Hence, LLMs can be chained together to create some kind of logic running through them. Rapid engineering, then, isn’t simply about trying to get them to do something extravagant with the LLMs. In our opinion, it is something more similar to natural language programming. It will need its own tools framework, similar to other forms of programming.

How big is the market?

TechCrunch+ explored why companies expect the enterprise generative AI market to grow to immense proportions. There should be many miners (customers) who will need picks and shovels (rapid engineering tools) to get the most out of Generative AI.

Vellum declined to share his pricing scheme, but noted that his services cost between three and four figures per month. Crossed over with more than three dozen clients, this gives Vellum a pretty healthy pace of execution for a start-up company. A rapid increase in demand tends to correlate with market size, so it is fair to say that there is indeed strong business demand for LLMs.

This is good news for the huge number of companies building, implementing or supporting LLMs. Given the number of startups in that mix, we were expecting bright, sunny days.