About
I build infrastructure for large-scale AI systems, with a focus on training…
Activity
671 followers
-
Gaurav Arora reposted thisGaurav Arora reposted thisBrother, can you spare a token? Prudence is important in our age of tokenmaxxing but that’s not the only reason we built Pinterest’s cloud-hosted VLM serving stack with NVIDIA Dynamo and Blackwell GPUs. We wanted a platform which could convert rich visual context to reusable embeddings to dramatically speed up real-time multi-modal AI while giving our teams a common foundation to ship new experiences like Pinterest Assistant, hybrid search and more. This was the result of close collaboration with NVIDIA and is also one of the very first use cases for PinCompute EKS, Pinterest’s next gen compute infrastructure. Many thanks to my co-authors, Lei Pan, Salina Wu, Cristian Lopez, Guantong Bai, Saurabh Vishwas Joshi, Chia-Wei Chen, Ambud S.en, Ambud S. and the incredible Pinterest and NVIDIA teams that made this possible! https://lnkd.in/gWNkXyeC cc: Elijah Soba, Qi Wang, Anthony Casagrande, Guan Luo, Kris(Yu-Hsin) Hung, Ryan McCormick, Harry Kim, Akshatha Kamath, Matt Rawson
-
Gaurav Arora shared thisHad a great time speaking at Ray Summit 2026 about The Synthetic Data Flywheel at Pinterest. We shared how we’re using Ray Data and vLLM to build large-scale synthetic data pipelines for post-training — combining GPU inference, tools, and LLM judges into scalable post-training workflows. Ray Data has been a great fit for these kinds of heterogeneous workloads, especially when dealing with stages that have very different scaling and concurrency characteristics. This work was truly a team effort. Huge thanks to Yash Upadhyay , Raphael Poulain, Tianjian Huang, Zhenyu Tan, Shubham Gupta, Shunyao Li, Luyan Wu, Korhan Citlak, Dave Chen, AI Training team at Pinterest! A special thank-you to Karthik A.ial thank-you to Karthik A. and Bo Liu for their leadership and support. Really enjoyed sharing what we’ve learned with the Ray community. Thanks to everyone who joined the session, to Eric W. for co-presenting this with me, and to Anyscale for putting together a great Ray Summit! #RaySummit #Ray #LLM #PostTraining
-
Gaurav Arora shared thisExcited to share our Pinterest Engineering Blog post on scaling distributed training for foundation recommendation models. A key challenge was embedding table sharding: the model was embedding-heavy, and cross-node All-to-All communication quickly became the bottleneck. By rethinking the sharding strategy with 2D sparse parallelism, we kept expensive embedding communication local within each node and used lighter-weight synchronization across nodes. This helped us reach near-linear scaling: 3.9x on 4 nodes and 7.5x on 8 nodes / 64 GPUs. https://lnkd.in/g_49xWkeAchieving Near-Linear Training Scalability for Pinterest’s Foundation ModelsAchieving Near-Linear Training Scalability for Pinterest’s Foundation Models
-
Gaurav Arora shared thisInteresting debugging by our group at Pinterest. Do read.Gaurav Arora shared thisAn ML stack upgrade (e.g., PyTorch, CUDA) always takes longer than expected, and we learn new things every time we undertake one. Chen Yang Chantat Eksombatchai Shunyao Li Mark Molinaro Shaochen Y. and Charles-A. Francisco captured some of our learnings this time, and there are some insightful tips on profiling and perf debugging. https://lnkd.in/g_yj5GKY
-
Gaurav Arora reposted thisGaurav Arora reposted thisWe're looking for an Engineering Manager to lead a team that transforms our ML Platform (training platforms, feature stores, dataset management, serving systems) into an end2end seamless experience. You'll bridge complex infrastructure with intuitive UX, driving the complete ML lifecycle transformation with ML Engineers at the center. We need someone with deep empathy for ML researchers, strong product sense, and experience leading teams that build developer-facing platforms. This is reimagining how ML Engineering gets done! Drop me a DM if interested.
-
Gaurav Arora reposted thisGaurav Arora reposted thisExcited about innovating cutting-edge website traffic management software with a focus on enhancing customer experience? Amazon-EdgeFabric is seeking a Software Development Engineer to contribute to our team! Explore a thrilling career opportunity at Amazon - EdgeFabric to embrace fresh challenges and shape the future together! 🚀 Job Opportunity: https://lnkd.in/gcNtZUpR #SoftwareDevelopment #AmazonSoftware Development Engineer, Traffic Engineering, Traffic EngineeringSoftware Development Engineer, Traffic Engineering, Traffic Engineering
-
Gaurav Arora reposted thisGaurav Arora reposted thisI'm hiring! Our team is working on developing cutting-edge website traffic management software that prioritizes customer experience. If you're passionate about creating top-tier software focused on networking, data processing, security, and utilizing AWS technologies, check out the job opportunity here: https://lnkd.in/g5fmhrSqSr. Software Development Engineer, Traffic Engineering, EdgeFabricSr. Software Development Engineer, Traffic Engineering, EdgeFabric
-
Gaurav Arora shared thisGaurav Arora shared thisThe selection process for our Senior Software Development Engineers consists of multiple interview rounds, aimed at gauging a candidate's capabilities with various skills required for the role. In the second episode of our Senior Engineer interview prep series, Alosh, Principal Engineer at Amazon, shares crucial tips and hacks to help you prepare on the fundamentals of Low Level Design. For more useful tips, keep an eye out for the next video in our series. Explore open Senior Engineering positions at Amazon: https://lnkd.in/g-eVBsF #InsideAmazonIndia #InterviewingAtAmazon #bepeculiar
-
Gaurav Arora liked thisGaurav Arora liked this𝗦𝗼𝗺𝗲𝘁𝗶𝗺𝗲𝘀, 𝘁𝗵𝗲 𝗯𝗶𝗴𝗴𝗲𝘀𝘁 𝗰𝗮𝗿𝗲𝗲𝗿 𝗱𝗲𝗰𝗶𝘀𝗶𝗼𝗻 𝗶𝘀 𝗰𝗵𝗼𝗼𝘀𝗶𝗻𝗴 𝘁𝗼 𝘀𝘁𝗮𝗿𝘁 𝗮𝗴𝗮𝗶𝗻. At the 𝗜𝗜𝗧 𝗛𝘆𝗱𝗲𝗿𝗮𝗯𝗮𝗱 𝗔𝗹𝘂𝗺𝗻𝗶–𝗦𝘁𝘂𝗱𝗲𝗻𝘁 𝗖𝗼𝗻𝗻𝗲𝗰𝘁 𝟮𝟬𝟮𝟲, held in 𝗦𝗲𝗽𝘁𝗲𝗺𝗯𝗲𝗿 𝟮𝟬𝟮𝟲, Kunal Aggarwal shared a career journey that took him far beyond the path he had originally chosen. After graduating in 𝗖𝗼𝗺𝗽𝘂𝘁𝗲𝗿 𝗦𝗰𝗶𝗲𝗻𝗰𝗲 𝗮𝗻𝗱 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗳𝗿𝗼𝗺 𝗜𝗜𝗧 𝗛𝘆𝗱𝗲𝗿𝗮𝗯𝗮𝗱 𝗶𝗻 𝟮𝟬𝟭𝟯, he began his career as a software developer at Goldman Sachs. Three years later, he made a very different choice. He left his job, prepared for the Civil Services Examination and joined the 𝗜𝗻𝗱𝗶𝗮𝗻 𝗥𝗲𝘃𝗲𝗻𝘂𝗲 𝗦𝗲𝗿𝘃𝗶𝗰𝗲 𝗶𝗻 𝟮𝟬𝟭𝟴. From software development to tax investigation and now international taxation, his journey is a reminder that 𝗼𝘂𝗿 𝗳𝗶𝗿𝘀𝘁 𝗰𝗮𝗿𝗲𝗲𝗿 𝗰𝗵𝗼𝗶𝗰𝗲 𝗱𝗼𝗲𝘀𝗻'𝘁 𝗵𝗮𝘃𝗲 𝘁𝗼 𝗱𝗲𝗳𝗶𝗻𝗲 𝗼𝘂𝗿 𝗲𝗻𝘁𝗶𝗿𝗲 𝗷𝗼𝘂𝗿𝗻𝗲𝘆. Sometimes, a change in direction is exactly what we need. But the conversation went beyond careers. 𝗔 𝗰𝗼𝗻𝘃𝗲𝗿𝘀𝗮𝘁𝗶𝗼𝗻 𝗮𝗯𝗼𝘂𝘁 𝘀𝘁𝘂𝗱𝗲𝗻𝘁 𝘄𝗲𝗹𝗹𝗯𝗲𝗶𝗻𝗴 Kunal also spoke about something that deserves much more attention: 𝘀𝘁𝘂𝗱𝗲𝗻𝘁 𝗺𝗲𝗻𝘁𝗮𝗹 𝗵𝗲𝗮𝗹𝘁𝗵. Academic pressure. Career uncertainty. Expectations. The fear of not being good enough. These are realities that many students experience, but don't always talk about. And as an alumni community, our responsibility should extend beyond career guidance. We need to build a community where students feel 𝘀𝘂𝗽𝗽𝗼𝗿𝘁𝗲𝗱, 𝗵𝗲𝗮𝗿𝗱 𝗮𝗻𝗱 𝗰𝗼𝗺𝗳𝗼𝗿𝘁𝗮𝗯𝗹𝗲 𝗮𝘀𝗸𝗶𝗻𝗴 𝗳𝗼𝗿 𝗵𝗲𝗹𝗽. Because success cannot only be measured by placements, salaries or professional achievements. 𝗧𝗵𝗲 𝘄𝗲𝗹𝗹𝗯𝗲𝗶𝗻𝗴 𝗼𝗳 𝗼𝘂𝗿 𝘀𝘁𝘂𝗱𝗲𝗻𝘁𝘀 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝗷𝘂𝘀𝘁 𝗮𝘀 𝗺𝘂𝗰𝗵. A big thank you to Kunal Aggarwal for sharing not only his career journey, but also a perspective that reminds us that taking care of ourselves and each other is an important part of building a successful life. Indian Institute of Technology Hyderabad/ Dean Alumni and Corporate Relations IIT Hyderabad / Rahul Garg / Kunal Aggarwal / Anand Konjengbam / Ashif Equbal, PhD / Amit Kumar Chaudhary / Dr. Abdul Mateen Ahmed #IITH #IITHAA #IITHAlumni #AlumniStudentConnect #IITAlumni #CareerJourney #PublicService #StudentWellbeing #MentalHealth #AlumniStories
-
Gaurav Arora liked thisOver the past year, our team has been working on JEV-like prefill-only inference in SGLang. Checkout what our team built:Gaurav Arora liked this🚀 New blog: Scaling JEV-like decision models with SGLang Decision models need a score, not prose. Classification, ranking, and agent action selection all ask the same thing: which option wins? Open-Jev, for example, scores each candidate separately with a Yes/No prompt. Serving this well raises two issues: Generate + top-k logprobs can drop the label you need, and the shared context can be recomputed for every candidate. SGLang addresses both: - /v1/score returns scores for the exact labels you request (Yes/No, A/B/C) - Multi-item scoring (MIS) computes the shared context once and keeps each candidate isolated - MIS latency stays nearly flat from 2 to 16 candidates, with 16-candidate p95 on Qwen3-8B dropping from 54.1 ms (Generate) to 20.6 ms (MIS) - MIS p95 stays under ~100 ms as load rises on Qwen3-0.6B, vs. seconds for Generate and SIS Huge thanks to the LinkedIn team for contributing! Benchmarks and launch commands in the blog: https://lnkd.in/gPj5U7TG
-
Gaurav Arora liked thisGaurav Arora liked thisHonored to be included in TIME's inaugural Executives of the Year: Tech & Data list. It's recognition of the awesome work Pinterest Engineering, Product, and Design teams and the cross-functional partners who build alongside us do every day. At Pinterest, we believe AI should make the product more useful without taking away the joy of discovery. Powered by our Pinterest Intelligence and our Navigator family of visual language models, we are turning intent into action —whether you’re discovering an idea, refining it, or finding something to buy. Thank you to TIME, and congratulations to all of this year’s honorees! https://lnkd.in/gnWj4GvZ
-
Gaurav Arora liked thisGaurav Arora liked thisThat's a wrap on the vLLM Conference at Ray Summit. Lines out the door for talks about an inference engine—that says everything about how much this community shows up. Thank you to Anyscale for co-hosting with us, and we're already looking forward to the next event. Every session recording is in the comments below:
-
Gaurav Arora liked thisGaurav Arora liked thisLast week at Pinterest Presents, our biggest event for advertisers, we shared how Pinterest Assistant is evolving to make discovery more visual, personal, and useful. It helps people find ideas, compare options, and move toward decisions with the taste and context that make their journey uniquely theirs. That experience is powered by Pinterest Intelligence: our first-party signals, visual understanding, recommendation systems, and fit-for-purpose AI models working together to anticipate what people want to try, buy or do next. We also announced Visual Search Ads, bringing that visual understanding to a new performance opportunity for brands as people search, explore and decide. It was a great event and I always enjoy meeting with our partners! The feedback on our consumer experience and ads products is always spot on and actionable. Read more and watch the Pinterest Presents sessions here: https://lnkd.in/eQY7iQTZ
-
Gaurav Arora liked thisGaurav Arora liked thisIn this latest engineering blog we shared some really cool work from our engineering team in partnership with NVIDIA where we describe key optimizations to achieve cost efficient VLM serving at Pinterest scale. https://lnkd.in/gskR_xxt One of the significant optimizations we did was for Pinterest Assistant where instead of feeding raw images into the VLM, we were able to reuse precomputed PinCLIP embeddings by running a projector that maps the embeddings into the target VLM’s native visual token space. This avoids the most expensive parts of pixel-based image serving. The performance impact of this approach has been significant. Compared with pixel-based image inputs in Dynamo, incorporating projection embeddings into Dynamo have yielded results that are substantially faster across our benchmarks: average gains are roughly 85x faster TTFT, 7.3x faster end-to-end latency, and 1.1x faster TPOT. Just as importantly, this allows us to have a much larger context supporting up to 25X larger visual context.Building Pinterest’s VLM Serving Stack on NVIDIA DynamoBuilding Pinterest’s VLM Serving Stack on NVIDIA Dynamo
Experience & Education
Licenses & Certifications
Projects
-
Recommendation System (B. Tech mini Project)
-
-
Mubase (mini-DBMS)
-
The project was development of database here @IITH, the database then supported creation/deletion of tables, predicated scan, sorting on relation’s attributes.
Implemented relation manager which adds and deletes the record, scans the records and handled predicated scanning on a particular relation.
Also did the merging with different layers of the database from bottom layer (record storing) to the above layer (sorting). Implementation is done in C++.
Other creators
Languages
-
English
Professional working proficiency
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