Suggestions
Raman SHRIVASTAVA
AI Engineer & Leader | LLMs, RAGs, AI Agents | 40 under 40 Data Scientist, AIM 2019
Raman Shrivastava is a Co-Founder and AI Lead at CustomerGlu, a gamified customer engagement and retention platform.13 He has 13 years of professional experience in Artificial Intelligence, both as an entrepreneur and in corporate roles.4 Shrivastava is recognized as one of the 40 under 40 Data Scientists by AnalyticsindiaMag in 2019.12
Prior to CustomerGlu, Shrivastava founded Fitsome, Inc., which was funded by Tim Draper and DFJ.1 He is an AI expert specializing in Large Language Models (LLMs), Retrieval-Augmented Generation (RAGs), and AI Agents.2
Shrivastava's educational background includes being an alumnus of VIT, Vellore, where he completed his BTech in Electronics and Instrumentation Engineering in 2011.6 He has also been involved in various AI initiatives, including leading advanced analytics and AI projects at INFINIT.5
At CustomerGlu, Shrivastava works alongside co-founders Prateek Gupta (CEO) and Sumant Subrahmanya (Product Lead) to develop innovative solutions for customer engagement and retention using AI and gamification techniques.13
Highlights
Github models: https://lnkd.in/gDTprGjF
key points:
- Models available: Llama 3.1, GPT-4o, GPT-4o mini, Phi 3, Mistral Large 2 etc.
- Playground(low/no code): access each model via a built-in playground that lets you test different prompts and model parameters in GitHub.
- Codespaces & VsCode integration: quick path to bring the models to developer environment in Codespaces and VS Code.
- Azure AI integration: easy integration for enterprise production grade deployments
Key points from the Nvidia H100 whitepaper
• H100 Tensor Cores: 6x AI performance boost vs A100 ○ New FP8 precision ○ Enhanced large language model processing
• FP8 precision advantages: ○ 2x throughput of FP16/BF16 ○ 50% smaller memory footprint ○ E4M3 and E5M2 formats
• Transformer Engine: ○ Dynamic FP8/FP16 precision switching ○ 9x faster AI training ○ 30x faster inference on large models
• First GPU with hardware-based Confidential Computing ○ Trusted Execution Environments (TEEs) ○ Enhanced data/code protection
• NVLink Network: ○ Connects up to 256 GPUs ○ 57.6 TB/sec all-to-all bandwidth
• DPX instructions: ○ 7x speedup for dynamic programming algorithms ○ Major boost for genomics processing
• Asynchronous execution improvements: ○ Tensor Memory Accelerator ○ Asynchronous Transaction Barrier
• HBM3 memory on H100 SXM5: ○ 3 TB/sec bandwidth ○ 2x increase over A100's HBM2
• Second-gen MIG technology: ○ 3x more compute per GPU Instance ○ 2x more memory bandwidth vs A100
• DGX H100 system: ○ 32 petaFLOPS of AI performance ○ Scalable to DGX SuperPOD configurations
