co-founder @raisi_ai | research collaborator on multi-agent simulations @ Berkeley | AI optimist
San Francisco
Joined March 2024
- Tweets829
- Following435
- Followers540
- Likes1K
We're excited to share that Raisi has been named a Top Ramp Partner for 2026! 🚀
Huge thank you to @aluatele and the @tryramp team for the recognition, support and belief in the future of what we're building together.
At Raisi, we bring together technology, relationships, and resources to help companies grow. Our partnership with Ramp extends that work through financial tools, shared insights, and events across our community.
The context layer directly affects the experience people get from AI.
All the tests I’m running now prove that, and the same was true a year ago. For example, in LongMemEval, GPT-4o scored 87% accuracy with relevant chat sessions and 60.6% with the full history.
I use @linear as the system of record for my agents when working on complex projects.
- Define what needs to be built, breaking it down into projects and issues with as much detail as possible.
- Add everything to Linear.
- Check the plan and sequence of work one more time to make sure they lead to the outcome you want, and look for anything missing.
- Give agents ownership of specific projects and have them post updates as they complete milestones.
- Have agents move work to "ready for review" when it's ready for your review.
This system is agent-agnostic, and I found this helps me even deliver on projects faster (agents are more productive because of it)
Dan Krieger retweeted
A few of our favorite moments from SF Tech Week.
From pitch preparation at Plug and Play to breakfast at Frontier Tower, deep-tech demos aboard the USS Hornet, and Founder Investor Day at Finix HQ, thank you to everyone who joined us to talk about building companies, raising capital, and what investors actually look for.
Thank you to our hosts, sponsors, and partners across the week: @PlugandPlayTC, Seed2Table, @frontiertower, @The_NetworkFirm, Contextful, @finix, @Navan, @BakerTillyUS, and @sprintoHQ.
A special thank you to our speakers and moderator for sharing their time, experience, and honest perspectives:
@RobertHarary, Dave Younts, Daven Nijran-Talwar, @KarnikaYashwant (KEY), Arseny Klekovkin, John Huang, Ashley Trick, @riohodges__, Brooke LeBlanc, Yash Alluri, Adam Henderson, Arjun Vad, @arjunraime, Kendrick Kho, @ldervisheva, @pedropiscodias, Pete Frischmann, Sarah Loveland, @1ankursethi, Adam Pluemer, Aidan Banafsheha, Vinit Bhansali, Rumika Sharma, Toly Kvitnitsky, Jiatao Fan, Amit Jain, @TweetsFromPawan, @thomastjoshi, Asit Kumar Sahoo, and @EmanuelPleitez.
To the founders who pitched, asked questions, and shared what they’re working on, and everyone behind the scenes who helped bring these events together: thank you.
See yourself in the recap? Share your favorite moment with us.
#SFTechWeek #Raisi
A team of 6 AI research agents spent 20 days building a Hearthstone bot that reached the top 100 on its server.
And the weird part is that the LLMs that built it couldn't play the game nearly as well themselves.
1/9
8/9
I think the bigger takeaway has very little to do with gaming.
An LLM doesn't necessarily need to be the best at doing a task itself.
It can build a system that's better at it, test it against reality, figure out what's wrong, and keep improving it.
That's what makes this interesting to me.
9/9
Here's the paper if you want to read it:
arxiv.org/abs/2610.33845
And the GitHub with the simulator and benchmark:
github.com/MCG-NJU/AutoTaver…
331 microsoft developers said what ai lets them do that they couldn't before. only 12% described new kinds of work. 45% described the same work, faster, and 37% work that used to be out of reach. the authors warn some could end up approving work they can no longer judge
self-reported survey at one company: about 4,000 developers were invited and 8.9% answered. the authors present approving work you can't judge as one of two possible paths, not a finding. paper: arxiv.org/abs/2610.23708
