Pavi (@_pavidhiman) header
Pavi (@_pavidhiman) profile picture
Joined September 2020
Photo by Pavi (@_pavidhiman)Photo by Pavi (@_pavidhiman)Photo by Pavi (@_pavidhiman)Photo by Pavi (@_pavidhiman)Photo by Pavi (@_pavidhiman)Photo by Pavi (@_pavidhiman)Photo by Pavi (@_pavidhiman)Photo by Pavi (@_pavidhiman)Photo by Pavi (@_pavidhiman)Photo by Pavi (@_pavidhiman)
the exciting part of autoresearch is how far an agent can take an idea with the right research environment. in inference, each improvement can reveal a new bottleneck and another direction to pursue thanks to the folks @DecagonAI who helped push this forward!
11
2
60
2,449
Pavi retweeted
inference serving is one of the cleanest targets for autoresearch. a lot of the attention right now is kernel gen, but the bigger surface is end to end. it’s a constrained optimization with a verifiable objective. hold latency and quality slas, maximize throughput. parallelism strategy, batching policy, cache config, speculator choice, routing, kernels, all in one search space. this optimization is also highly workload dependent. we’re cooking, more soon
8
5
1
155
8,260
RL agents need millions of real interactions to learn but for robotics, that's either extremely slow or expensive World models let agents learn by imagining trajectories and I wanted to understand how this actually works at the compute level So, I built DreamerV3 from scratch: autograd engine, RSSM, two-hot distributional critic, gaussian actor, fused GRU kernels, all without PyTorch The agent learns an internal physics model then trains a humanoid walking policy entirely in imagination with ~3h of training on a T4! The technical writeup covers the full architecture and some failed experiments: pavidhiman.com/dreamerv3.htm…
14
7
68
4,146
The agent's learned physics model imagining vs reality, this stays accurate for ~6s before drifting
4
419
I don’t think 9 year old Aryan thinks he’s met his goal yet. Excited to keep pushing @AIatDoorDash!
Today we are welcoming the Metis team to DoorDash as part of DoorDash AI Research. For the past six months, DoorDash has partnered with Metis to build AI agents together, and we have been consistently impressed by their team. By joining forces, we aim to accelerate our plans on building agentic commerce and pushing the frontier of physical intelligence. Excited to share more there soon. It’s still early innings with how AI will transform local commerce, and we’re looking forward to exploring those possibilities together with Aryan, Aayush, Marcus and the Metis team!
4
5
88
22,356
Super grateful for my journey with Metis and I’m excited to keep building as part of DoorDashs AI Research Team
Today we are welcoming the Metis team to DoorDash as part of DoorDash AI Research. For the past six months, DoorDash has partnered with Metis to build AI agents together, and we have been consistently impressed by their team. By joining forces, we aim to accelerate our plans on building agentic commerce and pushing the frontier of physical intelligence. Excited to share more there soon. It’s still early innings with how AI will transform local commerce, and we’re looking forward to exploring those possibilities together with Aryan, Aayush, Marcus and the Metis team!
5
13
1,273
note to self - zoom out & remember why you started
11
3
68
3,796
12/05/2025
14
27
3
342
13,907
Super grateful to be working with 2 incredible teams this summer: > ML Engineering Intern @PragmaClin – building tools for Parkinson’s patients > Robotics Engineering Intern @vanar_robots – building the next gen of affordable humanoids Excited to build in health and robotics!
2
14
1,326
brainstorming with a pen and a paper remains undefeated if you are trying to solve anything above the mildly complex difficulty
23
56
6
906
31,433
Got my first drone on my 9th birthday, have followed @zipline's drone design since I was 14 and built 2 drone projects within 2 months of turning 18. In short, I love drones. Check out my two projects below — hitchhiking drones for food delivery & 3D renders for disaster relief.
7
2
94
7,031
Hitch — enabling long-range & affordable delivery by letting drones hitchhike onto cars > Built a drone from scratch > Prototyped landing on moving targets: RL, ArUco, CV > Mapped fleet management, network effects & car choosing algorithm > Pitched to @CommaCapital & @AforeVC
1
14
1,143
Skycraft — viewing post-disaster zones through 3D models, generated from drone footage > Used Gaussian Splatting to convert drone footage into 3D renders in <1h > DHT11 sensor for real-time temp & humidity data > Web app with summary bot to analyze renders & sensor data together
1
14
1,511
2024 wrapped: > graduated high school > finished 1A of software eng @UWaterloo > became a @risefortheworld winner + Google scholar > continued building monitor.pd > learned to play poker! > joined @PragmaClin > built a drone from scratch > memories with old friends and new
7
1
53
4,175
Just released my newsletter covering: 1. Committing to university 2. Interning ⁦@PragmaClin⁩ 3. Building ⁦@HackTheNorth⁩ 4. Becoming a RISE Global Winner, Loran Finalist, etc. I hope you enjoy the read and feel free to reach out! open.substack.com/pub/pavidh…
3
493
Today on an all-new #TeenScientist, Rayna Malhotra talks with 16-year-old healthtech researcher @_pavidhiman about accessibility and AI in healthcare. 6 - 6:30 pm | 88.1 FM + wdiy.org
1
1
4
771
Monitor.PD won @_buildspace s4! The support has been unimaginable. To all my fellow final 32s, I loved every project - the content creators, gamers, musicians and the farmers This is the beginning, and I’d love any feedback thus far Thank you @farza, @ser_ando, @buildspace
11
115
11,426
congrats to the winners of s4 final 32.
48
50
18
512
59,140
I might win $25,000 today, but I need your help. My team and I are in @_buildspace's @_nightsweekends final 32. I usually don't ask for favors, but any and every vote counts. It would mean the world if you took 10s to vote!
2
8
1
43
5,536