- Joined Cognition
- Jul 2025
- How they joined
- With Windsurf
- Before Windsurf
- University of MarylandCMSC452 (Elementary Theory of Computation) TA
- Status
- Current
- Posted on X: "@ain3sh Congrats Ainesh. Let me know if you need help getting correct results from Devin CLI"Oct 30, 2026@ain3sh Congrats Ainesh. Let me know if you need help getting correct results from Devin CLI
- Posted on X: "FrontierCode 2.0 is here! With recent model releases, it almost feels like SWE is a solved problem. Yet anyone who uses coding agents regularly knows how frustrating it can be when models fail to think through details and deliver thorough implementations. Still more work to do!"Oct 27, 2026FrontierCode 2.0 is here! With recent model releases, it almost feels like SWE is a solved problem. Yet anyone who uses coding agents regularly knows how frustrating it can be when models fail to think through details and deliver thorough implementations. Still more work to do!
- Posted on X: "We worked closely with Artificial Analysis to bring FrontierCode to the Coding Index. I think it’ll be a great addition that helps us better understand model capabilities!"Oct 12, 2026We worked closely with Artificial Analysis to bring FrontierCode to the Coding Index. I think it’ll be a great addition that helps us better understand model capabilities!
- Posted on X: "Agree with this. Really happy to see that there’s finally a third party doing independent evaluation verification. This is really important for making sure the field doesn’t get stalled by benchmark quality issues, like what happened with SWE-Bench Pro when every model was stuck around 60%. I also really appreciate how much effort the Terminal-Bench authors have put into making sure we have a hig"Sep 19, 2026Agree with this. Really happy to see that there’s finally a third party doing independent evaluation verification. This is really important for making sure the field doesn’t get stalled by benchmark quality issues, like what happened with SWE-Bench Pro when every model was stuck around 60%. I also really appreciate how much effort the Terminal-Bench authors have put into making sure we have a high-quality benchmark. Even now, Terminal-Bench 2 and 2.1 are probably still my favorite benchmarks. They helped set the standard for what good coding benchmarks should look like. At the same time, I don’t think we should reduce every benchmark to a binary “good” or “bad” label and end up making people distrust everything out there. A third party should instead help set a clear standard for what makes a benchmark trustworthy, and work closely with eval providers to improve their methodology and produce evaluations that the whole field can actually rely on.
- Posted on X: "Since I started using Devin a lot late last year, I haven’t been able to go back. The interaction just feels natural, and it makes getting things done so easy. Do it all with Devin!"Aug 19, 2026Since I started using Devin a lot late last year, I haven’t been able to go back. The interaction just feels natural, and it makes getting things done so easy. Do it all with Devin!
- Public headline says "ML @ Cognition | ICPC ’23/24 World Finalist".
- Serves as student assistant coach for the University of Maryland ICPC team, with support role alongside coach Professor Mohammad Hajiaghayi.Mar 31, 2026
- Previous roles include Software Engineer Intern at Cognition; Software Engineer Intern at Windsurf; Technical Advisor Intern - GenAI at Scale AI; CMSC398L (Introduction to Competitive Programming) Instructor at University of Maryland; CMSC451 (Design and Analysis of Computer Algorithms) TA at University of Maryland; CMSC452 (Elementary Theory of Computation) TA at University of Maryland.
- The UMD team he assistant-coached (Danny Mittal, Sujay Konda, Austin Zou) earned a silver medal at the 2026 ICPC North America Championship.Mar 31, 2026
- Stated that there is still a lot more to coding than what any single benchmark can measure, and that this is another step toward a more complete picture.Oct 12, 2026
- The UMD team he assistant-coached advances to the ICPC World Finals in Dubai in November 2026.Mar 31, 2026
- Cheng-Yuan Lee competed in ICPC 2024 World Finals on 2024-09-19 with University of Maryland team UMD WHITE, ranked 134, score 2/12.Sep 19, 2024
Research: - LLM Evaluation - Data Pipeline - RL Framework - Inference With main focus on RL data and evaluations
On-policy RL infrastructure + Inference
(acquired by Cognition) Inference Infrastructure
Frontier Coding/Math/STEM, RL environments for SWE, Humanity's Last Exam
Fall '23, '24 / Spring '24, '25 Initiated the course & designed the syllabus.
Fall '23/'24
Spring '24
- Double Major in CS & Math - Two time ICPC World Finalist - ICPC Co-coach - President @ Competitive Programming Club - Algorithm Researcher - HSPC Problemsetter
@ain3sh Congrats Ainesh. Let me know if you need help getting correct results from Devin CLI
FrontierCode 2.0 is here! With recent model releases, it almost feels like SWE is a solved problem. Yet anyone who uses coding agents regularly knows how frustrating it can be when models fail to think through details and deliver thorough implementations. Still more work to do!
We worked closely with Artificial Analysis to bring FrontierCode to the Coding Index. I think it’ll be a great addition that helps us better understand model capabilities!
Agree with this. Really happy to see that there’s finally a third party doing independent evaluation verification. This is really important for making sure the field doesn’t get stalled by benchmark quality issues, like what happened with SWE-Bench Pro when every model was stuck around 60%. I also really appreciate how much effort the Terminal-Bench authors have put into making sure we have a high-quality benchmark. Even now, Terminal-Bench 2 and 2.1 are probably still my favorite benchmarks. They helped set the standard for what good coding benchmarks should look like. At the same time, I don’t think we should reduce every benchmark to a binary “good” or “bad” label and end up making people distrust everything out there. A third party should instead help set a clear standard for what makes a benchmark trustworthy, and work closely with eval providers to improve their methodology and produce evaluations that the whole field can actually rely on.
Since I started using Devin a lot late last year, I haven’t been able to go back. The interaction just feels natural, and it makes getting things done so easy. Do it all with Devin!
@ain3sh Congrats Ainesh. Let me know if you need help getting correct results from Devin CLI
FrontierCode 2.0 is here! With recent model releases, it almost feels like SWE is a solved problem. Yet anyone who uses coding agents regularly knows how frustrating it can be when models fail to think through details and deliver thorough implementations. Still more work to do!
We worked closely with Artificial Analysis to bring FrontierCode to the Coding Index. I think it’ll be a great addition that helps us better understand model capabilities!
Agree with this. Really happy to see that there’s finally a third party doing independent evaluation verification. This is really important for making sure the field doesn’t get stalled by benchmark quality issues, like what happened with SWE-Bench Pro when every model was stuck around 60%. I also really appreciate how much effort the Terminal-Bench authors have put into making sure we have a high-quality benchmark. Even now, Terminal-Bench 2 and 2.1 are probably still my favorite benchmarks. They helped set the standard for what good coding benchmarks should look like. At the same time, I don’t think we should reduce every benchmark to a binary “good” or “bad” label and end up making people distrust everything out there. A third party should instead help set a clear standard for what makes a benchmark trustworthy, and work closely with eval providers to improve their methodology and produce evaluations that the whole field can actually rely on.
Since I started using Devin a lot late last year, I haven’t been able to go back. The interaction just feels natural, and it makes getting things done so easy. Do it all with Devin!
