About
Activity
3K followers
-
Sumanth R Hegde shared thisExcited to share our work on large-scale sharded weight transfer! We’ve implemented a native sharded weight transfer engine in vLLM and SkyRL using Ray Direct Transfer (RDT) on NIXL. We’re able to achieve weight transfer for Kimi K2 (1T params) in BF16 in 7.53 seconds on 48 8xH100 nodes. There were two key problems to solve for sharded weight transfer: 1. How do you compute the shard to send to a vLLM worker on the trainer? 2. How do you scale to large models without blowing up memory and keeping transfer times low? For problem 1, the key issue is that the operations in weight loading can vary a LOT across models and layers. To support these, we use a “recording tensor” dry run - we use a dummy tensor to record the various transformations during weight loading, and store the op chain at initialization time. We replay the op chain on the trainer side to compute the full weights during each weight sync. For the second problem, we applied some tried and tested ideas: 1. PP-local and EP-local gather: Avoid redundant gather across PP groups, avoid gathering EP layers to avoid OOMs 2. Pipelined execution: Pipeline all-gather on the trainer, the weight transfer and the post-process on the inference side We have more details on the implementation as well as a demonstration of fault tolerance with RDT + NIXL in our blog. Check it out: https://lnkd.in/gEqAxwaU This work was led by Aaron Hao. It was great to have collaborated with the NVIDIA NIXL team for this effort - thanks to Gal Meirom, Istvan Haller, Gavin Parnaby, Moein Khazraee, Omri Kahalon who provided a number of useful tips for pushing performance further and ran the large scale benchmarks for Kimi K2.Large-Scale Sharded Weight Transfer with Ray Direct Transport (RDT) in vLLMLarge-Scale Sharded Weight Transfer with Ray Direct Transport (RDT) in vLLM
-
Sumanth R Hegde shared thisJoin us for the first-ever SkyRL Meetup! It's been a year since we released the SkyRL library. SkyRL is now one of the most popular open-source RL libraries with 2k+ GitHub stars and a growing user base in academia and industry, including researchers from Stanford, CMU, UC Berkeley, Microsoft AI, Datadog, and more. We're hosting a SkyRL Meetup on August 18th in SF. We have a special set of lightning talks from SkyRL users: 1. Chuck Tang, Trajectory - "Open Weights Aren't Enough: Frontier RL Training for Everyone" 2. Ben Cohen, Datadog - “Post-training an SRE Agent with SkyRL” 3. Eddie Richter, AMD - “Scaling Reinforcement Learning with SkyRL on AMD” 4. Charlie Ruan, University of California, Berkeley / Mercor - “Training Knowledge Work Agents with SkyRL” We’ll also do a deep dive into the library and talk about the road ahead. Registration + more details here: https://luma.com/y4ahouh6
-
Sumanth R Hegde reposted thisSumanth R Hegde reposted thisToday we are excited to announce, in partnership with the Google Kubernetes Engine (GKE) team at Google Cloud, a major milestone in Ray Serve LLM’s throughput and latency characteristics: Ray Serve LLM now matches high performance, rust-based routing frameworks such as vllm-router in benchmarks across a variety of workloads and deployment patterns. In our new blog, we cover three major optimizations to the Ray Serve LLM + vLLM stack that made this possible: direct streaming, a new vLLM Ray executor backend, and HAProxy integration. As a result, we see up to 4.4x higher request throughput than previous versions on prefill-heavy workloads, and up to 24x higher request throughput on decode-heavy workloads. Ray is a popular choice for complex, Python-native distributed computing batch inference pipelines with heterogeneous hardware. And now, we believe that Ray’s powerful primitives for fault tolerance, observability, flexibility across Kubernetes and VMs will enable the next generation of optimizations as LLM inference deployments become increasingly complex. Thanks to Spencer Peterson, Andrew Sy Kim, Kourosh Hakhamaneshi, Jeffrey (Yu-Che) Wang, Richard Liaw, Akshay Malik, Abrar Sheikh, and Alex Yang whose contributions to Ray Serve and Ray Serve LLM made this possible. A special thanks to the vllm-router (vLLM) and SGLang Model Gateway (SGLang) teams for great engineering on their respective projects. Read the full writeup here: https://lnkd.in/guHrz_FA
-
Sumanth R Hegde shared thisCheck out our work on native RL APIs for vLLM! Blog: https://lnkd.in/giq22pwPSumanth R Hegde shared thisExcited to share some of our work on improving vLLM for RL! A number of RL frameworks, including SkyRL, use vLLM for inference, and we’ve noticed some common problems: Weight syncing between training and inference is implemented in an ad-hoc fashion and duplicated across frameworks. Asynchronous RL is prone to break at scale, especially in P/D and DPEP deployments. We’ve been working on improving both! For more details check out: https://lnkd.in/geWNvSav 𝗪𝗲𝗶𝗴𝗵𝘁 𝘀𝘆𝗻𝗰𝗶𝗻𝗴 𝗶𝗻 𝘃𝗟𝗟𝗠 Weight syncing with vLLM has typically been implemented with ad-hoc worker extensions and RPC endpoints. While this works, it leads to a few issues. Most frameworks typically care about specifying the transport logic of how exactly vLLM will receive weights, but now they also need to deal with ad-hoc pre/postprocessing. Many frameworks also end up duplicating transport logic for popular strategies like NCCL, CUDA IPC as well as implementing the same optimizations (ex: packed tensor). This also leads to version locked implementations because they reach into vLLM internals. We introduce native APIs for weight transfer: /init_weight_transfer_engine, /start_weight_update , /update_weights and /finish_weight_update for differents stages of weight transfer. Along with these endpoints is a WeightTransferEngine abstraction allowing users to specify custom transport logic for receiving weights. We provide NCCL and CUDA IPC implementations out of the box, but framework developers can bring their own. The APIs being simple still allows for advanced use-cases like sharded weight transfer from M trainer ranks to N inference ranks. See my prototype here: https://lnkd.in/gbZAhNik 𝗔𝘀𝘆𝗻𝗰𝗵𝗿𝗼𝗻𝗼𝘂𝘀 𝗥𝗟 𝘄𝗶𝘁𝗵 𝘃𝗟𝗟𝗠 Async RL is the default for reasoning and agentic RL to maximize utilization with long tailed trajectories. We’ve worked on upgrading async RL with vLLM: - New pause mode to preserve requests in the scheduler: Users don’t need to manually bookkeep requests - Deadlock fixes for DPEP: This one took many iterations! DPEP requires careful coordination between vLLM engines for generation, and we ensure the same with weight syncing! The fixes have been tested at scale - Prime Intellect has validated async RL training for zai-org/GLM-5.1-FP8 in a P/D, DPEP32 deployment across 16 8xH200 nodes. It’s been great working with Sumanth R Hegde on this effort!
-
Sumanth R Hegde reposted thisSumanth R Hegde reposted thisNew research blog 🎉 Reinforcing Recursive Language Models Can a 4B model learn to recursively call itself and outperform frontier models on difficult long-context questions? We train a 4B-sized model to behave as a native Recursive Language Model. RLMs are a powerful new inference paradigm that lets a model recursively decompose problems, mitigating issues like context rot (the sharp decline in performance we feel when an LLM approaches large context sizes). RLMs have proven to be remarkably effective on long-context tasks, but can be tricky to get right with smaller models. We develop an RL framework to teach models to dispatch themselves onto subproblems and solve those subproblems effectively. On evidence selection across scientific papers, our 4B RLM matches Sonnet 4.6 in quality while running significantly faster and cheaper. While the blog presents an initial set of experiments, there's a lot more to explore with RLM training. We've added an RLM env to the SkyRL repo, check out the code and blog post on alphaXiv below! Thanks to Sumanth R Hegde and Charlie Ruan from the SkyRL team for helping resolve issues, and to Alex Zhang for feedback throughout the work!
-
Sumanth R Hegde shared thisWill be at the PyTorch Conference next week talking about SkyRL w/ Tyler Griggs ! Do join: https://lnkd.in/gnsw9Qd9Sumanth R Hegde shared this⚙️ Engineering for scale and speed. Next-gen RL libraries are redefining how agents learn and interact. 🎤 Open Agent Summit Speakers: • Sumanth R Hegde (Anyscale) – “SkyRL: A Modular RL Library for LLM Agents” • Tyler Griggs (UC Berkeley) – “Lessons in Agentic RL Modeling and Infrastructure” • Chenyang Zhao (SGLang) – “Optimizing Multi-Turn RL and High-Performance Inference in Agents” Explore reinforcement learning systems that bridge research and production for multi-agent intelligence. ✅ Register → https://lnkd.in/gidyFYZW 📋 Agenda → https://lnkd.in/g8-ACKJv #OpenAgentSummit #PyTorchCon #AgenticAI #RL #AgentPerformance
-
Sumanth R Hegde shared this𝗦𝗸𝘆𝗥𝗟 𝗻𝗼𝘄 𝗵𝗮𝘀 𝗻𝗮𝘁𝗶𝘃𝗲 𝘀𝘂𝗽𝗽𝗼𝗿𝘁 𝗳𝗼𝗿 ⚡ 𝗙𝗹𝗮𝘀𝗵𝗥𝗟! We now support Int8 and FP8 rollouts, enabling blazing fast inference - upto 1.7x for the DAPO recipe - without compromising performance! 𝗛𝗼𝘄 𝗱𝗼𝗲𝘀 𝗶𝘁 𝘄𝗼𝗿𝗸? Scalable RL training frameworks like SkyRL, VeRL, etc. use separate libraries/ frameworks for inference and training. Ex: vLLM for inference and PyTorch FSDP for training. Typically, computations in inference and training happen at half precision (BF16). FlashRL: https://lnkd.in/gd987Bpe (by Feng Yao, Lucas Liu et. al.) enables RL training with quantized rollouts, where inference is run at lower precision - Int8, FP8, etc. There are two key ingredients to make this work: 1. Truncated Important Sampling (TIS): Small differences in the sampling probability distribution (inference) and policy probability distribution can lead to instability during training. TIS ensures that the difference in rollout and policy probabilities doesn’t hurt performance. 2. Online quantization support: Policy model runs in BF16 while rollouts run in FP8 / Int8. FlashRL patches vLLM’s initialization and weight loading APIs to ensure compatibility. 𝗙𝗹𝗮𝘀𝗵𝗥𝗟 + 𝗦𝗸𝘆𝗥𝗟 𝗶𝗻𝘁𝗲𝗴𝗿𝗮𝘁𝗶𝗼𝗻 We’ve integrated both TIS and online quantization support in SkyRL. While the first bit is easy - this is a simple correction factor for your policy loss - the second part is tricky and requires careful handling of weight syncing at the trainer and the inference engine so that a) You don’t run into OOMs b) You don’t break CUDA Graph capture for vLLM. With these improvements, we support Int8 and FP8 rollouts for at-least 32B parameter models. Docs + examples: https://lnkd.in/gzqq2HTq
-
Sumanth R Hegde reposted thisSumanth R Hegde reposted thisWe’re looking forward to participating in the Agentic AI Summit 2025 in Berkeley, CA on August 2nd. Find us onsite to speak to experts in PyTorch & vLLM. Plus don’t miss: ➡️ Focus talk on how The Linux Foundation and PyTorch Foundation are accelerating open source AI and strategies on how to keep agentic AI open. ➡️ Panel discussion on Frameworks & Stacks for Agentic Systems with Moderator: Chi Wang and Panelists: Ion Stoica, Matei Zaharia, Sherwin Wu, Chi Wang, Jerry Liu, Matt White, Bradley Axen ➡️ Workshop on Architecting an Open Agentic Future with Open Source and Open Standards featuring: 🔹 "Open Source and Open Standards: Foundations for an Agent-Driven Future" by Matt White, Executive Director, PyTorch Foundation, GM of AI, The Linux Foundation 🔹 "Accelerating Agents: Scalable and Efficient Inference with vLLM" by Michael Goin, Principal Engineer, Red Hat 🔹 "MCP as a Foundational Protocol for Agents" by Jason Kim, Member of Technical Staff, Anthropic 🔹 "Ray: Scalable Agent Infrastructure" by Sumanth R Hegde, Software Engineer, Anyscale 🔹 "Training Agents with Reinforcement Learning" by Daniel Han, Co-Founder & CEO, Unsloth AI 🔹 "A2A: The Future of AI Agent Collaboration" by Chitra Venkatesh, Product Manager, Google #OpenSourceAI #AgenticAI #OpenSource #AI #PyTorch #vLLM
-
Sumanth R Hegde reposted thisSumanth R Hegde reposted this📢 If you're in SF Bay Area on Thursday night, come to our event for drinks and discussion about post-training of AI agents. We'll have folks from Veris AI, Anyscale and the SkyRL project among others. It is going to be a lot of fun. Thanks to Rho for sponsoring. Registration link is in the comments.
-
Sumanth R Hegde liked thisSumanth R Hegde liked thisAs more teams scale open source AI on secure, governed infrastructure, Ray has become critical across the full AI lifecycle, from data curation to production serving and reinforcement learning. Ahead of PyTorch Conference North America 2026 in San Jose, we put together a guide highlighting key Ray-focused sessions on the schedule. Featured speakers include: Anyscale / Databricks / University of California, Berkeley: Ion Stoica Anyscale: Eric Tang, Sumanth R Hegde, Joshua Lee, Mengjin Yan Google: Ankita Luthra, Kotturu Trinadh, Jago Macleod LinkedIn: Tao Huang, Tommy Li Pinterest: Gaurav Arora, Shunyao Li, Eric W. Uber: Ke Chen, Peng Zhang, Xandra Zhu University of Minnesota: Arun Sharma Explore engineering takeaways from these technical sessions where you will learn about unified AI orchestration with Kubernetes, elastic training stacks, and scalable RL. Read the full guide and map out your schedule: https://lnkd.in/e373G-Hu Register for PyTorch Conference North America: https://hubs.la/Q04tBgv_0 #PyTorchCon
-
Sumanth R Hegde liked thisSumanth R Hegde liked thisToday we're announcing that Frank Slootman is joining Nscale's Board of Directors and Adam Rosman is joining as our Chief Legal Officer. We are operating at a scale, and moving at a speed, that puts a premium on experience. On people who have built at this size before and know how to get it right. Frank has led and scaled some of the most consequential enterprise technology companies of the last two decades. Adam has spent 25 years steering companies through the legal and regulatory demands of growth at this pace. We get the benefit of all of that, and I'm looking forward to it. Welcome, Frank. Welcome, Adam. https://lnkd.in/gimYu9kFNscale appoints Frank Slootman to the Board of Directors and Adam Rosman as Chief Legal Officer | NscaleNscale appoints Frank Slootman to the Board of Directors and Adam Rosman as Chief Legal Officer | Nscale
-
Sumanth R Hegde liked thisSumanth R Hegde liked thisAMD ALERT🚨🚨 On Kimi K3 2.8T, AMD has even better profit margins per GigaWatt Compared to B300, GB200 NVL72, B200. This is due to AMD's lower TCO and better performance leading to better profit margins of 53.6% on Mi355X compared to 44.3% on GB200 NVL72. This is from our open source Agentic Inference benchmark, AgentX Shoutout to Ramine Roane & Emad Barsoum 10x engineers for optimizing this 2.8T model!
-
Sumanth R Hegde liked thisSumanth R Hegde liked thisFidji Simo is joining the Nscale Board of Directors. Fidji has built products used by billions of people. She led the Facebook app through the industry's shift to mobile, took Instacart to profitability and through its IPO as CEO and Chair, and most recently ran ChatGPT as CEO, AGI Deployment at OpenAI. Compute allocation is a product problem as much as a power problem, and that gets more true the further up the stack we build. Very few leaders understand what products at that scale demand of the systems underneath them. Fidji has spent her career on the other side of it. We've built this board deliberately. Fidji is the piece we didn't have. Welcome, Fidji. https://lnkd.in/gzEDBjvUFidji Simo Joins Nscale Board of Directors | NscaleFidji Simo Joins Nscale Board of Directors | Nscale
-
Sumanth R Hegde liked thisSumanth R Hegde liked thisWorked with Mercor Research and the SkyRL team, training Qwen3.5-397B-A17B on APEX-Agents (long-horizon office work) off-the-shelf data with SkyRL, improving Pass@1 from 16% to 27%. The post is more of a practical field guide for what to do given an RL dataset, de-risking step by step: environment infra, harness fixes/optimizations, token-in-token-out, an overfitting run, small-model ablations, then the hero run. We were also delighted to see the gains transfer to OpenCode despite training on an MCP-based harness, and even to Terminal-Bench 2.1. Code, models, and eval traces are all open source. Hope it helps your next hillclimb! Blog: https://lnkd.in/gbcaZiiP Source code: https://lnkd.in/g5XkSeYX
-
Sumanth R Hegde reacted on thisSumanth R Hegde reacted on thisA higher KV cache hit rate doesn’t always mean lower TTFT and TPOT, or higher throughput. KV cache reuse is only one part of the routing story. Balancing token load across replicas is more important, especially for heterogeneous workloads such as asynchronous RL rollouts with stragglers and agentic workloads. Token load is simply the total work a replica needs to handle: the incoming request’s remaining prefill work after accounting for KV cache reuse, plus the active prefill and decode work already on that replica. This gives us a more accurate picture of how much work each replica actually has to do. In Ray Serve LLM, we introduce a router that considers both KV cache affinity and token load when deciding where to route each request. As a result, we see roughly 15% improvement in TTFT and TPOT over the session-affinity baseline when replaying coding agent traces. In our new blog, we explore the tradeoff between KV cache reuse and token load balancing. We also show why traditional microservices routing approaches like consistent hashing, while effective at maintaining a high KV cache hit rate for multi-turn conversations, are not granular enough to balance highly heterogeneous LLM workloads. Thanks to Seiji Eicher and Kourosh Hakhamaneshi for the inspiring discussions, and special thanks to Rudy Pei and the NVIDIA Dynamo team for modularizing Dynamo’s components like KV indexer for Ray Serve LLM to integrate with for request scoring. Finally, I’ll present this work at the vLLM conference Aug 26 2 pm. Hope to see you there! Read the full writeup here 👉 https://lnkd.in/gksGKJQfOptimizing LLM Serving Efficiency: Moving Beyond KV Cache Reuse to Token-Load Awareness with Ray Serve LLM | AnyscaleOptimizing LLM Serving Efficiency: Moving Beyond KV Cache Reuse to Token-Load Awareness with Ray Serve LLM | Anyscale
-
Sumanth R Hegde reacted on thisSumanth R Hegde reacted on thisExcited to share that our paper “Data Intelligence Agents: Interpreting, Modeling, and Querying Enterprise Data via Autonomous Coding Agents” has been accepted to the EMNLP 2026 Industry Track! Now it is Official! DIA uses autonomous coding agents to take on the full data intelligence workflow, from understanding raw enterprise data, to building semantic models, to answering complex questions through executable, validated artifacts. Across seven SQL benchmarks, DIA outperforms the best prior systems on 6 of 7, including substantial gains on some of the hardest benchmarks. A great milestone for the team at C3 AI, and more evidence for something we strongly believe: autonomous coding agents can be a powerful abstraction far beyond software development. Huge congratulations to Anoushka Vyas, Aarushi Dhanuka and Henrik Ohlsson #EMNLP2026 #AutonomousAgents #CodingAgents #AgenticAI #DataIntelligence #EnterpriseAI
-
Sumanth R Hegde liked thisSumanth R Hegde liked thisEveryone knows that RDMA is fast. But how do you actually realize the speedups in your own application? As we were developing Ray Direct Transport (RDT), we found that there a lot of nuances in how Python applications use RDMA libraries like NVIDIA's NIXL, to the point where just 𝘂𝘀𝗶𝗻𝗴 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 𝗹𝗶𝗯𝗿𝗮𝗿𝘆 𝗰𝗮𝗹𝗹𝘀 𝗰𝗮𝗻 𝗺𝗮𝗸𝗲 𝗮 7.5𝘅 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗶𝗻 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲. Our goal with RDT was to make RDMA easy and seamless to use for Ray applications like weight syncing for RL for LLMs. To do that, we also had to make sure performance was easy to achieve. Check out our blog post that walks through how to do this yourself using RDT. These steps are the key to the speedups that RDT achieves for weight syncing in end frameworks (𝘮𝘰𝘳𝘦 𝘰𝘯 𝘵𝘩𝘦𝘴𝘦 𝘪𝘯𝘵𝘦𝘨𝘳𝘢𝘵𝘪𝘰𝘯𝘴 𝘴𝘰𝘰𝘯): • 3.5s for Qwen3-235B-A22B with SkyRL - 18x faster than NCCL • 2.3s for GLM-4.5-Air with Miles - 2.2x faster than NCCL and 1.14x faster than Mooncake Blog link: https://lnkd.in/gCQSf-gz Joint work with my awesome teammates at Anyscale: Joshua Lee Xinyu Zhang Sumanth R Hegde Aaron Hao (and former teammate Dhyey Shah). Thanks also for feedback from Mengjin Yan Richard Liaw.
Experience & Education
Volunteer Experience
-
Coding Events Volunteer
Shaastra, IIT Madras
- 1 month
Science and Technology
-
Informals Volunteer
Saarang
- 1 month
Science and Technology
-
Education Blog Volunteer
National Service Scheme
- 11 months
Social Services
Part of the IIT Madras chapter of the National Service Scheme. Led the Education Blog project, aimed at educating the uninitiated in science and math.
Publications
-
Accelerating Direct Preference Optimization with Prefix Sharing
2024 Conference on Neural Information Processing Systems, Fine-Tuning in Machine Learning Workshop
See publicationOffline paired preference optimization algorithms have become a popular approach for fine-tuning on preference data, outperforming traditional supervised fine-tuning in various tasks. However, traditional implementations often involve redundant computations, especially for tasks with long shared prompts. We introduce prefix sharing for preference tuning, a novel technique that processes chosen and rejected responses as one sequence with a shared prefix. To prevent cross-response contamination…
Offline paired preference optimization algorithms have become a popular approach for fine-tuning on preference data, outperforming traditional supervised fine-tuning in various tasks. However, traditional implementations often involve redundant computations, especially for tasks with long shared prompts. We introduce prefix sharing for preference tuning, a novel technique that processes chosen and rejected responses as one sequence with a shared prefix. To prevent cross-response contamination, we use a custom block-sparse attention mask. Our method achieves 1.1-1.5× improvement in training throughput on popular DPO datasets, without any effect on convergence. When combined with sequence packing, we observe consistent 1.3-1.6× speedups, benefiting even datasets with smaller sequence lengths. While we focus on Direct Preference Optimization (DPO), our approach is applicable to other paired preference tuning methods. By enhancing computational efficiency, our work contributes to making preference-based fine-tuning more accessible for a wider range of applications and model sizes.
-
Image Restoration for Under-Display Cameras
European Conference on Computer Vision (ECCV) 2020
See publicationDeveloped a novel two-stage deep network for restoration of Under-Display Camera images. The network first restores image quality at a lower resolution followed by joint upsampling. The two-stage setup enables fast reconstruction while also producing sharp, high-quality images.
My Presentation: https://youtu.be/WnNOg178iSk
Courses
-
Deep Generative Models
CSE 291B
-
Design and Analysis of Algorithms
CSE 202
-
Graduate Networked Systems
CSE 224
-
Natural Language Processing
CSE 256
-
Object Oriented Programming
CS2810
-
Operating Systems
CSE 120
-
Pattern Recognition and Machine Learning
CS5691
-
Principles of Database Systems
CSE 232
-
Recommender Systems and Web Mining
CSE 258
-
Scalable Data Systems
DSC 204a
Projects
-
Parameter Efficient Fine Tuning
-
See project- Worked on parameter-efficient fine-tuning(PEFT) methods for large language models.
- Implemented and benchmarked (IA)^3, a new state-of-the-art method for efficient fine-tuning, supporting various decoder-only and encoder-decoder models (such as GPT, OPT, RoBerta, etc), 8-bit quantization, multiple adapters, etc.
- Our code serves as the official implementation for (IA)^3 in the HuggingFace PEFT library -
Operating System Kernel Internals
-
- Built different parts of a UNIX-like operating system in C, implementing multi-threading, process scheduling and synchronization
- Implemented a user-level thread package with thread creation, scheduling logic, etc and process scheduling functionality, supporting different policies like round robin, stride scheduling, etc -
Surfstore: A Dropbox-Like Service
-
- Built a scalable, distributed DropBox-like cloud storage service for file syncing, with fault-tolerance support.
- Developed a MetaStore service for storing file metadata and a BlockStore service for storing file blocks using gRPC.
- Implemented the RAFT consensus protocol and employed consistent hashing for horizontal scaling. -
Jester: A Text-to-Meme Generation Engine
-
See project- Built a novel two-stage system to generate relevant meme templates and meme captions given any user text.
- Implemented a flexible softmax-free transformer model to serve candidate meme templates for given user text, achieving a top-5 accuracy of 71\% on a dataset of 300,000 captions.
- Utilized GPT-3 and designed custom prompts for 100 templates to generate relevant meme captions from user text.
Poster:…- Built a novel two-stage system to generate relevant meme templates and meme captions given any user text.
- Implemented a flexible softmax-free transformer model to serve candidate meme templates for given user text, achieving a top-5 accuracy of 71\% on a dataset of 300,000 captions.
- Utilized GPT-3 and designed custom prompts for 100 templates to generate relevant meme captions from user text.
Poster: https://drive.google.com/file/d/1dWeZCb53pRt3-um7uRC2lrkU-KBpsxUq/view?usp=sharing -
Low Resource Satellite Image Segmentation
-
– Achieved image segmentation of satellite images into 8 classes with an average accuracy of 96.27%, and placed 4th in Inter IIT Tech Meet 2018.
– Executed a mix of classical computer vision and deep learning methods with only 14 images for training.
– Implemented U-net based architectures and employed hard mining for under-represented classes.
Honors & Awards
-
President of India prize
Indian Institute of Technology Madras
Awarded the President of India prize for the best academic performance among all 2021 graduates of IIT Madras
-
Guest, Republic Day Parade 2020
Ministry of Human Resource Development, Government of India
Invited to the Republic Day Parade as a guest of the Hon'ble Prime Minister of India
-
Kishore Vaigyanik Protsahan Yojana scholarship
Department of Science and Technology, Government of India
Awarded the KVPY scholarship to pursue research in the basic sciences, along with a provisional admission into the Indian Institute of Science.
Test Scores
-
Graduate Record Examinations
Score: 339/340
Verbal - 169/170
Quantitative - 170/170
Writing - 5.0/6 -
BITSAT
Score: 450/450
Achieved a score of 450/450 on the BITSAT admissions test
-
JEE Advanced 2017
Score: 279/366
Secured AIR 504 , among more than 200,000 applicants
-
Karnataka Common Entrance Test KCET 2017
Score: 177/180
Secured 2nd rank in the state among more than 100,000 applicants
-
JEE Mains 2017
Score: 320/360
Achieved an All India rank of 93, among 1,200,000+ applicants
Languages
-
English
Professional working proficiency
-
Hindi
Limited working proficiency
-
Kannada
Native or bilingual proficiency
Organizations
-
Computer Vision and Intelligence Club
Stategist
-- Guided and led a team of 30 undergraduates working on cutting-edge technology. - Presented workshops for 150+ audience of IIT Madras students on computer vision and machine learning. - Mentored club projects in pattern recognition, autonomous mapping & navigation and reinforcement agents.
-
Evolve, Shaastra 2019
Organiser
-- Organised a Sports Technology Summit as a part of the annual technical fest of IIT Madras. - Spearheaded a Sports Analytics Hackathon, where 200+ developers brainstormed on problem statements employing machine learning. - Collaborated with Global Sports Commerce to host panel discussions, hackathons and fireside chats – Designed problem statements for AI applications such as player performance, consumer experience
Other similar profiles
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content