Yaateh is our Founding Machine Learning Engineer.
His advice for companies building an AI agent:
Don't give it everything.
Most companies we talk to shove in every document and data point they can find, thinking more context always leads to better results.
But "everything" tends to include the spreadsheet nobody has kept up to date, and the same numbers sitting in a few different reports that don't quite match.
In one 2025 study, an AI model got 84% of questions right when given the document containing the correct answer.
Add one more document that looked relevant but didn't contain the answer, and accuracy dropped to 73%.
"That level of accuracy loss is just not acceptable at scale" , according to Yaateh.
He's actively testing whether newer models are more resilient to this.
Until that's proven, he believes companies need to be deliberate about what they let their agents read, and thinks the people closest to the work should be the ones who check it.
Once the person who owns a process signs off on how it's written down, that's the strongest signal you have of what's correct.
So if you're building an agent, always ask:
How much of what your agent reads has been verified by someone who knows the work best?
Lisa is our Co-Founder and CEO.
Her advice for companies looking to start implementing AI:
Find the person inside your company who's most excited about it.
Because people get more comfortable with AI when someone next to them is already using it.
In Lisa's words, having people like that around "gives other employees the confidence to try it out and fail."
And the funny thing is, you usually don’t have to "appoint" this person.
They tend to reveal themselves.
The people who are most excited about AI usually can’t help but experiment with it, talk about it, and share what they’re learning with everyone around them.
That person will become your AI champion naturally.
Lisa’s simple advice for finding them:
“Ask yourself, who was the first person at your company to ask for a Claude subscription?”
The @TalvyHQ team is made up of best friends from college.
When our CEO, @KjHardrict, was an investor at @Link_Ventures, he had one core belief:
Always bet on founders who build with their best friends.
Now he's on the other side of that bet, and so are the companies we work with.
Most of them are full of people who've worked together for years too, and usually their biggest worry about AI is whether it will be the reason someone they've loved to work with has to be let go.
But our goal isn't to take away anyone's work, other than the parts that they never wanted to do in the first place.
Because the relationships that make companies special are not something you can replace with technology.
The person who has been there for a decade, who understands the company better than anyone, who has helped build everything from the ground up, is usually the person you want to make more capable, not replace.
That's the future we're trying to build toward.
In the same way KJ bet on founders who chose to build with their best friends, we believe the best teams should get to keep building with their favorite people.
Explaining lines from our Times Square Billboard:
“AI keeps getting better. Is your company?”
Every few months, the models get more capable.
But a better model doesn’t automatically make a company better at using it.
We meet companies that want agents capable of answering questions about their business, but the information those agents need is still scattered across different systems, buried in reports someone has to download manually, or isn’t being captured consistently in the first place.
The next model won’t fix that.
There’s foundational work companies can do today to make each new generation of AI more useful to them.
Getting important information into places AI can reliably access. Rethinking workflows that were designed around people doing everything manually. Teaching teams how to work differently as the technology improves.
AI is going to keep getting better either way.
The question is whether your company is getting better with it.
Talvy retweeted
builders don't understand how true this is.
people no longer want new AI tools to help them get work done, they simply.... want the work DONE lol.
companies are tired of buying products, they just want results delivered at AI-speed.
big reason why we built @TalvyHQ 😎
Team Talvy was in Silicon Valley this week for the @imaginationxyz summit.
Our CEO @KjHardrict spoke on a panel called "What Runs Without You," covering the problems our team is seeing inside companies trying to get automated, which usually has to do with a lack of ownership.
Most companies don't have an assigned person whose sole job is to choose what work goes to a model, write down exactly how that process gets done, and answer for it when it breaks.
Until someone owns that, nothing will ever truly run without you.
Our guys got a lot out of these two days.
Special thanks to @johnkwerner and Theo Ehrmann for putting the event together and having us out!
we love seeing our team in rooms where everyone thinks AI can make the world a better place.
thanks @imaginationxyz for hosting our boys!
a few things i'm noticing while at the Silicon Valley event today with the @TalvyHQ team:
- people are opening their imaginations to bigger and bigger problems with AI.
- there really isn't a societal problem in this room that somebody doesn't think AI could tackle.
- there are plenty of people with positive outlooks on AI’s future. the doomers are not everywhere!
Our CEO and COO met because a college housing office put them in the same room.
Perhaps that’s how we learned to treat client engagements like roommates, not houseguests.
A houseguest sees the version of your house you cleaned for them.
A roommate sees the sink full of dishes.
Most AI consultants act like houseguests at other companies, and they build for perfect conditions.
Our engineers build for everyday life.
By moving in alongside your people, we see how things look when they’re going according to plan.
But more importantly, we see what prevents the plan from always working:
The manual workarounds, the buggy old software people feel forced to use, and the little inefficiencies people have dealt with for so long they’ve stopped noticing them.
We see how the sink gets full of dishes.
And instead of helping you wash them faster, we figure out why they keep piling up.
Because you learn a lot more about a company by living with the mess than by being shown around the house.
We hope we can bring this feeling to everyone.
AI should be used to inspire others to reach their creative potential and remove the parts of the job they enjoy the least.
do more of the work you love.
AI is the reason my Co-Founder started coding again.
Lisa has something I call “Ultimate Taste.”
She just has this natural feel for how something should work before she can necessarily explain why.
But even though she has a CS degree from Harvard, spent time as an FDE at Palantir, using that superpower meant spending most of her week on the engineering part she liked least:
Coding.
So when she joined Talvy, she made it VERY clear what she didn’t want to spend time on.
But then one day I showed her Claude Code.
And her reaction was:
“Oh…. I can just ask it to do something?
This is basically what I would tell an engineer.”
Since realizing this, she’s back behind an IDE with a newfound excitement for building.
But instead of writing most of the code herself, she describes what she wants to the AI, looks at what comes back, says what’s off, and prompts again.
Her whole day basically runs on taste.
And I don’t think Lisa’s experience will be unique.
I think this can become a much more common way people work if they embrace AI the right way.
For most of history, having great taste wasn’t enough.
You also needed the technical ability, time, or resources to turn what you could imagine into something real.
AI is starting to remove that prerequisite.
The distance between “I know what this should be” and “I can make this exist” is collapsing.
And I think a lot of people are about to discover they’re much more creative than they thought.
Because now they can build at the speed of their own taste.