Phagwara, Punjab, India | LinkedIn | vipulpower2009@gmail.com

I build high-throughput distributed training stacks and ultra-low-latency serving infrastructure for frontier foundation models. My work bridges the gap between deep learning research and high-performance hardware, specializing in maximizing Model Flops Utilization (MFU) and optimizing GPU memory access at the metal level.

• 🛠️ Current Focus: Deep training optimizations (NVFP4, Muon, ZeRO++) and custom Triton/CUDA kernel compilation.
• ⚡ Superpower: Profiling bottlenecks at the warp level using Nsight Systems/Compute and writing fused kernels to keep GPUs saturated at line rate.
• 🏢 Previous Work: Senior AI Systems & Infrastructure Engineer at NiftyBooks (Austria-based startup), where I lifted LLM serving throughput ~4x and cut VRAM ~4x (NVFP4 = quarter the bytes on memory-bound decode) at p99 <200ms, and ran multi-GPU B200 fine-tuning at scale.

Python

C++

CUDA / Triton

PyTorch

Linux / Bash

Docker

• Distributed Training: Megatron-LM (TP/PP/DP), DeepSpeed ZeRO-1/2/3, ZeRO++ (hpZ, qgZ), PyTorch FSDP, Slurm.
• Serving & Inference: vLLM v1/v2, PegaFlow (Rust KV-cache daemon), TurboQuant (KV compression), P-EAGLE Speculative Decoding.
• Hardware Architecture: Qualcomm Snapdragon 8 Gen 2 / Elite (NPU/DSP), Jetson Orin, NVIDIA Blackwell B200 / Hopper H100.
• Analysis: Nsight Compute, Nsight Systems, PyTorch Profiler, KernelBench.

A high-performance pipeline simulating native NVFP4 mixed-precision pretraining on Blackwell 5th-gen Tensor Cores.

• Micro-block Scaling: Implemented dual-level scaling (16-element blocks with per-block FP8 scale + global FP32 scale) to limit outlier quantization error.
• Stochastic Rounding: Integrated stochastic rounding and random Hadamard transforms to prevent gradient accumulation bias and activation energy collapse.
• Triton Attention: Optimised a custom fused attention kernel utilizing Blackwell's 256KB Tensor Memory (TMEM) per SM, hitting 71% of peak B200 FLOPS (1.8x throughput increase).
• Auto-Tuning: Plugs into an autonomous reinforcement learning tuning loop to optimize hyperparameters (QK-clipping, Newton-Schulz iterations) without human intervention.

An offline, high-efficiency speech transcription system for code-mixed Indian languages running on edge hardware.

• Indic Curation: Fine-tuned AI4Bharat IndicConformer (120M) on Hindi-English code-mixed speech with telemetry-grade noise profiles.
• Quantization: Compressed model from 480MB to 145MB via INT8 dynamic quantization with <3% accuracy decay.
• Latency & Power: Achieved sub-80ms first-word latency on Snapdragon 8 Gen 2 with NPU/DSP acceleration.
• Dual Decoding: Hybrid CTC+RNNT engine for fast batch transcription (CTC) and real-time streaming input (RNNT).

Gymnasium environment training a reinforcement learning policy for stable quadcopter hover under wind disturbances.

• Trained a continuous PPO agent in a 3D simulation with 12-DOF physics (RK4 integration) to control motor thrust curves.
• Generalized to unseen wind vectors (0-5 m/s) with <3cm steady-state position error.

• 2.7x speedup on custom Triton attention kernels over baseline compiled PyTorch code.
• 60% drop in production serving costs (₹18L $\rightarrow$ ₹7.2L/mo) via disaggregated serving, PegaFlow cache, and TurboQuant.
• 0% checkpoint corruption over multi-day pretraining runs via SHA-256 dual-write verification.
• 94.7% intent preservation on Indic ASR under noisy conversational conditions.