Rahul Manish Jauhari

Hello, I am Rahul, currently pursuing MS in Computer Science at Northeastern University, Boston. I'm interested in AI inference, exploring intersections of AI and hardware.

At the moment, I am:

seeking full-time opportunities where I can apply my skills.
reading and swimming.

Work Experience

Software Development Engineer at AMD: spent 2 years working on ZenDNN, researching and developing hardware-aware optimizations for deep learning inference on AMD EPYC CPUs, spanning optimized kernels, graph execution, quantization, and large-scale AI workloads.

In my previous experience

designed and built a multi-instance vLLM benchmarking stack (Skill: N instances behind an nginx load balancer, a shared Hugging Face cache, offline-by-default instances).
built a repeatable vLLM + zentorch (ZenDNN backend) install path (Skill: pip CPU wheel or source with optional PR cherry-picks, version-gated against zentorch's supported range, local ZenDNN checkout, native or Docker, plus smoke and benchmark validation).
root-caused a vLLM + zentorch regression that broke multi-instance memory behavior via a hardcoded env var.
ran competitive analysis against Intel Granite Rapids, Turin and Venice, filing gaps as tracked defects.
built and owned the TF-Java interface for ZenDNN's TensorFlow plugin from nothing to a shipping product — benchmarks, accuracy infra, a user guide, an open-sourced examples repo on Hugging Face, and a public blog post.
upstreamed PluggableDevice support into the TensorFlow-plugin project, so zentf can be consumed from stock TF rather than a fork.
run the version migration treadmill — TF-Java 0.4.0 → 1.1.0, TensorFlow 2.18 → 2.20 → 2.21 — each with build fixes, revalidation, and accuracy regression hunts.
shipped op and fusion work: FusedMatMul (MatMul + BiasAdd + Tanh), biasadd+fusion across Conv2D and MatMul, ZenSafeEmbeddingLookup fusion, and ZenDNN-accelerated GatherNd/GatherV2.
added Conv3D support via an im2col + MatMul rewrite, driven by reading the TransNetV2 paper to derive the requirement.
migrated the plugin off the legacy ZenDNN library onto ZenDNNL, including the direct MatMul API and FEMM integration.
built and iterated the team's performance dashboards — daily/weekly views, zentorch-vs-IPEX comparisons, a data-ops admin UI — and reviewed roughly 130 pull requests across the ZenAI repositories.
stood up ZenDNN ONNX Runtime (onnxrt) benchmarking across Linux and Windows early on, the source of most of the team's automated CI reporting.
supported external engagements with Red Hat/RHAII, Snapchat, Netflix, BharatGen, and Byblos Bank on CPU sizing and benchmarking.