Indraneel Ghosh

About Indraneel Ghosh

At AWS, I work on all aspects of model performance and reliability of the EC2 UltraServer GPU fleets. This involves optimising TFLOPs for different distributed training and inference workloads to develop highly optimised recipes. These recipes are used to detect cluster performance issues and drive initiatives with internal and external stakeholders to maintain a high high fleet utilisation while providing the lowest fleet failure rates across all cloud providers. At Google, I worked on all aspects of provisioning highly performant, scalable and reliable GPU/TPU compute clusters on Google Kubernetes Engine. These efforts were critical to GCPs strategy for becoming a top-tier AI cloud.At Amazon, I worked on driving monetization initiatives for Prime Video that helped maintain a 20% ARR growth rate. I was also a founding engineer for Amazon MXPlayer where I built systems that scaled the offering from 0 to 7M+ DAU. I led projects related to video streaming, networking performance optimisation, content and ad recommendations.If you are a startup building in the AI infra space, feel free to drop a message to schedule a quick chat.Please don\'t reach out for referrals at Amazon.Open to interesting AI infra/full-stack/backend opportunities. Current Areas of Interest: Distributed Model Training and Inference Infrastructure, Model Performance Benchmarking, KubernetesPrevious/Other Areas of Interest: Video Streaming, Recommendation Systems, Computer Vision

Experience

1. Software Engineer

   Amazon Web Services

   Jul 2025 — Present · Seattle, WA, US

   AWS NitroModel Performance and Ultraserver Reliability- Scoped out charter for all aspects of Ultraserver fleet quality measurement and utilisation. This involved working with hardware engineering and platform teams across AWS to define quality standards for the GPU Ultraserver product- Re-architected existing workflows to a simplified self-healing system to reduce the operational load associated with Ultraservers. This helped reduce the manual effort involved for the team by 90% and grew our fleet utilization rate to 83%- Built measurement systems to evaluate and improve fleet monitoring and qualification systems. This involved building a new quality evaluation and testing framework that evaluates test efficacy and tunes hyperparameters for various tests to define stable configurations- Led efforts related to improving delivery velocity and fleet quality for NVL72 GB300 ROCE clusters. This involved working with multiple different organisations across AWS including Hardware Engineering, Data Center Operations, Annapurna and Nitro Networking- Working on Ultraserver performance benchmarking, monitoring, and optimisation initiatives(Platforms: GB200, GB300).Virtual Private Cloud- Optimised ENI allocation and improved noisy neighbor detection and mitigation algorithms. This effort involved securely profiling and improving the memory allocation of the Nitro hypervisor and defining metrics to identify noisy neighbor scenarios reliably- Implemented custom memory allocation algorithms(TCMalloc, jemalloc) to optimise performance and cost of running C/C++ systems that manage the VPC network allocation logic and operate at xx million TPS. The release led to an 8% TCO reduction.

Education

• Birla Institute of Technology and Science, Pilani

  Bachelor of Engineering - BE, Computer Science

  2021

• Pace Junior Science College, Nerul

  Senior Secondary School, PCM with Electrical Maintenance

• Apeejay School, Nerul

  Secondary School