- Joined
- May 2026
- How they joined
- Directly
- Before Cognition
- AmazonResearch, Amazon AGI SF Lab
- Status
- Current
- On SWE-2 launch he wrote verbatim: "SWE-2 is our strongest model yet and our first with different reasoning efforts.
We’ve show that careful RL design can advance the Pareto of cost and intelligence.
On FrontierCode, we’re within one point of fable while 64% cheaper. https://t.co/MWiujb0y17"Sep 10, 2026SWE-2 is our strongest model yet and our first with different reasoning efforts.
We’ve show that careful RL design can advance the Pareto of cost and intelligence.
On FrontierCode, we’re within one point of fable while 64% cheaper. https://t.co/MWiujb0y17
- On Cognition's fundraise he wrote verbatim: "Cognition’s run-rate revenue has grown from $492M to almost $900M since May -- and we're just getting started. Today's fundraise allows us to be more ambitious, bring exceptional people on, and scale our bets further. Come join us!"Sep 8, 2026Cognition’s run-rate revenue has grown from $492M to almost $900M since May -- and we're just getting started. Today's fundraise allows us to be more ambitious, bring exceptional people on, and scale our bets further. Come join us!
- Listed as a co-author of Cognition's 'Introducing FrontierCode' coding eval (June 2026) alongside Eric Lu, Ben Pan, Sam Lee, Ray Wang, Rohan Choudhury, Fermi Ma, TC Qin, Carlo Baronio and S. Alberti, and of 'FrontierCode 1.1' (July 7, 2026); his core Cognition research team is visible from these bylines.Jun 2026
- Joined Cognition in Research in May 2026 after ~2 years at Amazon AGI SF Lab (Jul 2024-May 2026); the timing puts him there just before the SWE-2 and FrontierCode pushes he now posts about.May 2026
- Listed as an author of Cognition's 'SWE-1.7: Frontier Intelligence at a Fraction of the Cost' (Ben Pan*, Carlo Baronio*, Rohan Choudhury, Eric Lu, Ryan Kim, Deniz Birlikci, TC Qin, ...).2026
- Personal site (written while at Amazon): he led reinforcement learning at Amazon AGI Labs on the Nova Act browser agent, helped take it 'from zero to a full AWS service' powering Alexa+ for 1M+ users, and 'built out and led the RL scaling team'; earlier he 'ran multimodal post-training for our research preview'.Dec 2025
- Current role is Research at Cognition.
- On SWE-2 demand he wrote verbatim: "SWE-2 has shown us every second counts when you ship a model people love. We were able to secure additional compute in record time, and now, we're making SWE-2 free in Devin Cloud until Oct 8th. It's a sweet (but cautionary) tale that emphasizes just how important compute planning is in this age. Compute is scarce and deals often take time. I'm very glad we've been able to move fast and make this happen."Sep 18, 2026SWE-2 has shown us every second counts when you ship a model people love. We were able to secure additional compute in record time, and now, we're making SWE-2 free in Devin Cloud until Oct 8th. It's a sweet (but cautionary) tale that emphasizes just how important compute plann
- On SWE-2 cost efficiency he wrote verbatim: "I think this is our craziest result. SWE-2 is ~80% cheaper than SWE-1.7 model despite being 3x larger. It shows that a larger model can be more cost-efficient per task, even if it's more expensive per token, because it runs more efficiently. Astra's cheaper than Sol. Fable 5.1 is cheaper than Opus. Now SWE-2 is cheaper than SWE-1.7."Sep 10, 2026I think this is our craziest result. SWE-2 is ~80% cheaper than SWE-1.7 model despite being 3x larger. It shows that a larger model can be more cost-efficient per task, even if it's more expensive per token, because it runs more efficiently. Astra's cheaper than Sol. Fable 5.1
- Recorded Episode 4 of Amazon's 'Making a Mind' podcast, 'Improving Agent Reliability with Reinforcement Learning', with host Danielle Perszyk, covering the RL stage for Nova Act (environments, verifiers, RL infrastructure); the episode frames 'Reliability, not accuracy' as the bottleneck for web agents.Feb 4, 2026
- LinkedIn post on the Nova Act launch, verbatim: "What a joy it is to be part of the team that made Nova Act happen! ... It's been a huge research & storytelling milestone for me. I got to both: 1. learn from amazing researchers and train our model 2. design a vision for our launch and bring it to life"; he owned launch storytelling as well as training.Mar 31, 2025
- Joined Adept as part of its incoming new-grad class in July 2024 and moved to Amazon with the Adept acquihire (Amazon hired Adept's founders), which is how he landed at Amazon AGI SF Lab.Jul 2024
zero to one for Nova Act, Amazon's browser use agent led the RL scaling team
Part of Adept’s incoming new grad class. Moved to Amazon with the acquihire.
Building multimodal AI search for fashion e-commerce.
Researching interactions between gesture and speech. Published my first paper into ACM.
Shared my passion of programming with 1000+ other students. Lead the creation of the experimental version of the course (git branch) and made it the norm in three semesters (git merge).
Head Teaching Assistant for 15-112, Swartz Center for Entrepreneurship Innovation Scholar, Sigma Phi Epsilon Swartz Center for Entrepreneurship Innovation Scholar: One of 13 students chosen for a 2-year entrepreneurial program. Includes a proseminar in entrepreneurial leadership, a Silicon Valley Trek with grants, a stipend for own startups, and a robust mentorship program.
Optimizing inference is now much easier with agents -- making it a no-brainer for high-TPM orgs. Decagon used Devin to do so. When I met @_pavidhiman and @cyrusasg a month ago, it was clear how killer Decagon research team was. Expect to see much more bangers from them soon.
Being able to share painful lessons openly makes the leaders that we look up to... human! In some ways, knowing that the people look up to are not that different can help you take bigger swings.
SWE-2 has shown us every second counts when you ship a model people love. We were able to secure additional compute in record time, and now, we're making SWE-2 free in Devin Cloud until Oct 8th. It's a sweet (but cautionary) tale that emphasizes just how important compute plann
Devin Fusion now uses SWE-2 to give you the same performance as Fable 5.1 Max at 36% less cost. There's a lot of juice to squeeze both on research and the harness side. Cognition is in a unique position being able to do both.
SWE-2 is our strongest model yet and our first with different reasoning efforts. We’ve show that careful RL design can advance the Pareto of cost and intelligence. On FrontierCode, we’re within one point of fable while 64% cheaper. https://t.co/MWiujb0y17
I think this is our craziest result. SWE-2 is ~80% cheaper than SWE-1.7 model despite being 3x larger. It shows that a larger model can be more cost-efficient per task, even if it's more expensive per token, because it runs more efficiently. Astra's cheaper than Sol. Fable 5.1
Cognition’s run-rate revenue has grown from $492M to almost $900M since May -- and we're just getting started. Today's fundraise allows us to be more ambitious, bring exceptional people on, and scale our bets further. Come join us!
I'm convinced scope violations are the biggest UX problem in coding models right now. It's remarkable how much frontier models fail to write simple implementations, and crazier how much headache it causes. We were early w/ FrontierCode; I expect more evals to capture this soon
We love Devin and want you to be able to run it inside any execution environment.
Optimizing inference is now much easier with agents -- making it a no-brainer for high-TPM orgs. Decagon used Devin to do so. When I met @_pavidhiman and @cyrusasg a month ago, it was clear how killer Decagon research team was. Expect to see much more bangers from them soon.
Being able to share painful lessons openly makes the leaders that we look up to... human! In some ways, knowing that the people look up to are not that different can help you take bigger swings.
SWE-2 has shown us every second counts when you ship a model people love. We were able to secure additional compute in record time, and now, we're making SWE-2 free in Devin Cloud until Oct 8th. It's a sweet (but cautionary) tale that emphasizes just how important compute planning is in this age. Compute is scarce and deals often take time. I'm very glad we've been able to move fast and make this happen.
