In partnership with

Today’s startup is a first for PreSeed Now. They’re not technically a UK startup, but their British founders are split between Seattle and London and they’re building something it’s worth taking a deep dive on.

What’s more, Yara co-founder Joe Braidwood is someone I’ve had on my radar for well over a decade, as you’ll see as you read on…

But first:

• Interested in quantum tech? As part of my work with Infinity in the Netherlands, I’ve co-authored a detailed report on quantum use cases in the energy sector. Take a look!

– Martin

Free Notion and Unlimited AI

Thousands of startups use Notion as a connected workspace to create and share docs, take notes, manage projects, and organize knowledge—all in one place. We’re offering 3 months of new Plus plans + unlimited AI (worth up to $3,000)! To redeem the Notion for Startups offer:

1. Submit an application using our custom link and select Beehiiv on the partner list.

2. Include our partner key, STARTUP4110P67801.

Learn More

Yara is harnessing AI to make mental wellness more accessible. Is AI up to the task?

Yesterday, I opened up a chat box in my browser and spoke to my new therapist online. 

Yara was an empathetic listener, helping me unpick my problems through targeted questioning, reinforcing statements, and breaking my problems down into manageable chunks to discuss.

As you will probably have guessed, Yara is an artificial therapist built on generative AI. And the reason I’m particularly interested in this one is that Yara is co-founded by Joe Braidwood who was a key figure in a UK startup I used to write about years ago. 

SwiftKey was the best of a wave of predictive typing keyboard apps that emerged in the early days of iOS and Android app stores. After it was acquired by Microsoft, which still offers the app on Android, Braidwood went on to develop his career on the west coast of the USA.

And now, based in Seattle, he’s teamed up with London-based Richard Stott to build a therapy app that really helps show how text-focused AI has moved on from the days when predicting the next word you would want to type was a novelty.

Braidwood says Yara has been built to take advantage of memory and token capabilities in current generative AI models.

“You can start to build more nuanced frameworks that can understand different factors of human psychology. What you end up with as an MVP is something that is just deeply attentive, that shows tremendous empathic responses and that paves the way for a new, very accessible daily habit in getting therapeutic advice.”

In its initial form, Yara, is being built for users who, Braidwood says, “are motivated to make changes in their lives, who recognise patterns, but perhaps don't quite have the strategies for coping with those patterns.

“It can be an always-on presence in their lives, here for them any time, where they can come back and they can continue to work on these things in their life that they want to improve.”

Is AI suitable for therapy?

The idea of using generative AI to make therapists available to more people is one that enthusiasts of the tech will embrace, but surely the inaccuracies and hallucinations the tech is known for make it a risky thing for Yara to offer?

Stott says the way the AI is built is key here.

“At its core, we have got a non-deterministic set of algorithms which are operating, but you can put checks and balances into those systems. You can make agents which will look at a draft response and say, ‘that's not kosher’, ‘that's not going to fly’, or ‘that's not safe’, or ‘actually, we need to sort of err on the side of caution here’. There are all sorts of layers of technology we can put in to make it safer.”

Stott adds that evaluating the quality of Yara’s output is critical to keeping it safe.

“Where there are things that fall short, we iterate, and we make it better and safer and higher quality, and that's the way that anything should happen in the scientific realm. Being ethical and being responsible will be at the core of our operation.”

Braidwood says Yara has started to assemble an AI safety team to look at using smaller, purpose-trained models.

“We're not there yet in publishing our thesis on this, but certainly the trend is giving us a lot of promise around not just building tremendous amounts of scaffolding around monolithic models, but actually really starting to construct more purpose-built architectures that are intrinsically safe and that have been heavily validated.”

What about privacy?

The other issue with AI, or indeed any internet-based service, is privacy - especially when you’re meant to be sharing your innermost thoughts with Yara.

“There's an interesting balance between harnessing the potential of AI and the requirement to move data from your mind into your device, and into the cloud, where these models are powered,” says Braidwood.

“And so we've been very considered about this. Everything is encrypted, both in transit and at rest. We have industry-standard, quantum-proof, passkey-ready tools to keep all of our data that we do store with our cloud partners incredibly under wraps.

“I also think that there's a potential future where some of this data can stay on device, as we get better at inference, as we get better at purpose-trained models. We're not there yet, but it's a really important point, and it's something that is fundamental in our thinking.”

The story so far

Braidwood says he first developed an interest in mental health when he was studying at Cambridge and shadowed his mum working in mental health in West London.

“One of the things I worked on was a web-based project to improve communication to the local community about mental health challenges. And I just could not believe what I was reading. This was a massive crisis. There were all these tools that existed and no-one was using them.

“And then I went and wrote a paper about that while I was studying policy at Brown, and started to do back of an envelope mathematics about the true compounding cost to employers around productivity and everything else.”

Maintaining an interest in mental health through his career in tech, it was recent developments in the sophistication of generative AI and a conversation over coffee with Stott that prompted the pair to team up for the transatlantic effort that is Yara.

Stott has a background in mathematics and clinical psychology and shares Braidwood’s interest in improving access to mental health support.

Yara began in private alpha with trusted testers, and then expanded to a larger vetted group of users before it recently opened as public beta you can try on the startup’s website.

The tech underpinning the chatbot continues to be updated. Braidwood says that last week they introduced an update to its ‘clinical brain’. 

“It combines the state of the art in a number of different ways of thinking about semantic representation and memory, agentic frameworks and supervision, and some of the safety guardrails that we've already discussed.”

Braidwood says users’ experiences so far point to an AI therapist being a useful complement to a human therapist, as well as an alternative to one.

“We’ve heard from people that have been in therapy for a long time, but keep pushing up against their innate reputational constraints, and so don't fully reveal quite what's going on, but are very happy to go into that detail with Yara. I think that's a really interesting, alternative way of thinking about the dynamic, which has been quite surprising and quite exciting.”

The plan now is to iterate on the current MVP until they find product-market fit. Along this path. They want to get to a point where users make Yara a daily habit and partnerships are in place to get it to the people who need it.

As for business models, the startup is exploring options for premium consumer tiers; offering Yara to businesses as an employee benefit, and finally Braidwood and Stott believe there’s a role for Yara as a licensed healthcare therapy.

Subscribe to Premium to read the rest.