# Ramakrishnan Sivakumar

> Principal ML Software Engineer at AMD working on on-device AI: running large language models efficiently on consumer hardware. Co-creator and maintainer of Lemonade, an open-source SDK for local LLM inference.

Ramakrishnan Sivakumar is a Principal ML Software Engineer at AMD in Seattle focused on on-device AI: running large language models efficiently on consumer hardware. Key facts: co-creator and maintainer of the Lemonade SDK (open-source local AI server, 5k+ GitHub stars); co-creator of TurnkeyML and GroqFlow; previously at Groq and Intel; two granted US patents; publications on AI accelerator evaluation. Expertise: On-device AI, Local LLM inference, Large language models, Machine learning systems, ML compilers, AI accelerators, AI accelerator benchmarking, Model quantization, Neural processing units (NPU), AMD Ryzen AI, ONNX, ONNX Runtime.

## Profile

- [Resume (Markdown)](https://ramkrishna.dev/resume.md): full machine-readable resume
- [Home](https://ramkrishna.dev/): bio, experience timeline, current focus, selected projects
- [LinkedIn](https://www.linkedin.com/in/ramakrishnansivakumar/): ramakrishnansivakumar
- [GitHub](https://github.com/ramkrishna2910): ramkrishna2910
- [Medium](https://medium.com/@ramkrishna2910): @ramkrishna2910
- [Google Scholar](https://scholar.google.com/citations?user=15TuaaoAAAAJ): Ramakrishnan Sivakumar
- [X](https://twitter.com/ramkrishna2910): @ramkrishna2910

## Current Focus

- [Lemonade](https://github.com/lemonade-sdk/lemonade): Co-creator & maintainer. Open-source local AI server: LLMs, speech, and image generation on your own hardware, with NPU and GPU acceleration.
- Tiered intelligence: Building. Getting every request to the right tier of compute - NPU, GPU, or cloud - through policy-driven model routing: rule, classifier, semantic, and LLM-as-router.
- Upstream open source: Contributor. Ongoing contributions across the local-AI stack: vLLM backends, llama.cpp enablement, and the ONNX ecosystem.

## Projects

- [Lemonade](https://github.com/lemonade-sdk/lemonade): Co-creator & maintainer. Open-source local AI server - LLMs, speech, and image generation on your own hardware
- [TurnkeyML](https://github.com/onnx/turnkeyml): Co-creator. ONNX toolchain for evaluating and accelerating models across hardware
- [GroqFlow](https://github.com/groq/groqflow): Co-creator. One-function-call compilation from PyTorch/ONNX to the Groq LPU

## Writing

- [Day 0 Support for Gemma 4 on AMD Processors and GPUs](https://www.amd.com/en/developer/resources/technical-articles/2026/day-0-support-for-gemma-4-on-amd-processors-and-gpus.html): How AMD delivered day-0 support for Google's Gemma 4 across Ryzen AI processors and Radeon GPUs, so the model runs locally through Lemonade the day it ships.
- [Lemonade by AMD: A Unified API for Local AI Developers](https://www.amd.com/en/developer/resources/technical-articles/2026/lemonade-for-local-ai.html): The flagship introduction to Lemonade as AMD's answer to fragmented local AI: one API surface over heterogeneous engines and hardware, from Ryzen AI NPUs to Radeon GPUs.
- [Ryzen AI and Radeon are ready to run LLMs Locally with Lemonade Software](https://www.amd.com/en/developer/resources/technical-articles/2025/ryzen-ai-radeon-llms-with-lemonade.html): A practical guide to running large language models locally on AMD hardware: Lemonade as the serving platform that picks the best available acceleration, from NPU to integrated and discrete GPUs, behind the OpenAI API standard.
- [Run OpenAI's gpt-oss locally with Lemonade](https://lemonade-server.ai/news/gpt-oss.html): Day-0 guide to running OpenAI's open-weight gpt-oss models fully locally with Lemonade - co-authored with the Lemonade maintainer team.
- [Rethinking Local AI: Lemonade Server's Python Advantage](https://www.amd.com/en/developer/resources/technical-articles/2025/rethinking-local-ai-lemonade-servers-python-advantage.html): Why Lemonade Server is built in Python and what that unlocks: rapid engine integration, a hackable stack for the community, and no performance penalty where it matters - the inference engines underneath.
- [Local Tiny Agents: MCP Agents on Ryzen AI with Lemonade Server](https://www.amd.com/en/developer/resources/technical-articles/2025/local-tiny-agents--mcp-agents-on-ryzen-ai-with-lemonade-server.html): Running Model Context Protocol agents entirely locally: Hugging Face Tiny Agents driven by NPU-accelerated LLMs on Ryzen AI through Lemonade Server - agentic AI without the cloud.
- [Unlocking a Wave of LLM Apps on Ryzen AI Through Lemonade Server](https://www.amd.com/en/developer/resources/technical-articles/unlocking-a-wave-of-llm-apps-on-ryzen-ai-through-lemonade-server.html): How an OpenAI-compatible local server turns the existing ecosystem of LLM applications - chat UIs, coding assistants, automation tools - into NPU-accelerated local apps on Ryzen AI, without code changes.
- [Why do companies Open Source Software?](https://ramkrishna2910.medium.com/why-do-companies-open-source-software-fc1b2c508206): Seven distinct motivations that drive companies to open-source software - from commoditizing complements to building hiring pipelines - and how to read a corporate open-source move for what it actually is.
- [GraphCast: A breakthrough in Weather Forecasting](https://ramkrishna2910.medium.com/graphcast-a-breakthrough-in-weather-forecasting-d70fae9ac365): How DeepMind's GraphCast uses graph neural networks to produce 10-day global weather forecasts that outperform traditional numerical weather prediction - and what it signals about ML displacing physics-based simulation.
- [BTS: The Marketing Masters in Disguise](https://ramkrishna2910.medium.com/bts-the-marketing-masters-in-disguise-54dea60492e4): What the world's biggest band can teach product builders about community, distribution, and brand loyalty - a look at BTS's playbook through a marketer's lens.
- [The Power of Product Ecosystems](https://medium.com/design-bootcamp/the-power-of-product-ecosystems-3f6afd41ca51): Why platforms win through ecosystems rather than individual products: a look at how developer tools, integrations, and community compound - with lessons for ML hardware companies building software moats around silicon.
- [OpenAI's ChatGPT introduces itself](https://medium.com/@ramkrishna2910/openais-chatgpt-introduces-itself-f612e92d9be5): Written in the week ChatGPT launched: an early hands-on exploration of what conversational LLMs could do, letting the model describe its own capabilities and limitations.
- [What, Why and How ONNX Script?](https://medium.com/@ramkrishna2910/what-why-and-how-onnx-script-74dd21ab396f): ONNX Script lets developers author ONNX functions and models directly in Python with eager-mode evaluation - dramatically improving debuggability over hand-writing graph protos. What it is, why it matters for the ONNX ecosystem, and how to use it.
- [Developer Velocity You Can't Unsee: Introducing GroqFlow](https://groq.com/developer-velocity-you-cant-unsee-introducing-groqflow/): The launch post for GroqFlow: why compiling a model to a novel AI accelerator should be one function call, not a week of toolchain wrangling - and how GroqFlow turns PyTorch and ONNX models into Groq LPU executables automatically.
- [The Promise of ONNX](https://medium.com/mlearning-ai/the-promise-of-onnx-d43c371f913d): Why an open interchange format matters when every ML framework has its own graph representation: ONNX as the interoperability layer that lets models move between training frameworks, inference runtimes, and hardware accelerators.
- [Optimizing BERT for Intel CPUs using ONNX Runtime default execution provider](https://cloudblogs.microsoft.com/opensource/2021/03/01/optimizing-bert-model-for-intel-cpu-cores-using-onnx-runtime-default-execution-provider/): How to accelerate BERT transformer inference on Intel CPU cores using ONNX Runtime's default execution provider - covering graph optimizations, quantization, and the practical export path from framework model to optimized ONNX, published in collaboration with Microsoft's ONNX Runtime team.

## Talks

- [Panel: Agent Infrastructure Roundtable 2026](https://luma.com/9k0g4dqj): Aramas AI & Berkeley Gateway Accelerator, Berkeley, CA, 2026. Panelist alongside leaders from Stanford, Salesforce, and the MBZUAI Institute for Foundation Models, moderated by Intel: what changes in the infrastructure stack when agents move from prototypes to persistent, production-scale systems, and where silicon, governance, memory, and orchestration need to land.
- [Vibe Coding with Local Models](https://www.amd.com/en/corporate/events/advancing-ai/sessions-catalog/vibe-coding-with-local-models.html): AMD Advancing AI 2026, San Francisco & Shanghai, 2026. Hands-on workshop delivered at AMD's flagship Advancing AI event in San Francisco (with Iswarya Alex) and at Advancing AI Day in Shanghai: harnessing the on-device power of the Ryzen AI Max 395 to run large coding models locally with Lemonade Server, no cloud, no API keys. Attendees architect specialized AI agents with Qwen Coder and watch them collaborate to build something they'll actually want to play.
- [Build with local AI agents using OpenCode and Lemonade](https://build.microsoft.com/en-US/sessions/LABSP586-R1): Microsoft Build 2026, San Francisco, 2026. Hands-on lab at Microsoft Build: powering real development workflows with local AI infrastructure, building and running applications with AI-driven coding agents, OpenCode as the agent and Lemonade serving the models on-device.
- [Lemonade running local LLMs on AMD - CES 2026 booth demo](https://www.youtube.com/watch?v=-fscysymnWY): CES 2026, AMD booth, 2026. Live demo at the AMD booth at CES 2026: Lemonade serving large language models fully locally on AMD hardware, showing NPU- and GPU-accelerated generation on consumer machines.
- [Panel: Interoperability and Standardization in AI](https://community.linuxfoundation.org/events/details/lfhq-lf-ai-data-foundation-presents-onnx-community-meet-up-1/): ONNX Community Meetup 2025, Intel campus, Santa Clara, 2025. Moderated the closing panel of the ONNX Community Meetup 2025 on interoperability and standardization across the AI stack.
- [ONNX and TurnkeyML - AMD engineer interview](https://www.youtube.com/watch?v=iG6z5eaTD24): AMD Developer Central, 2024. A conversation with AMD engineers Victoria Godsoe and Krishna Sivakumar on ONNX, the TurnkeyML toolchain, and how the ONNX Model Zoo was rebuilt on top of it.
- [The AI Revenue Problem](https://open.spotify.com/episode/7wwkopVh05U6zcybmfcySu): The Unenlightened Podcast, 2024. On the large gap between AI capital expenditure and revenue, why customer-facing AI products matter, and a brainstorm of AI product ideas with market and societal relevance.
- [The AI Revolution](https://open.spotify.com/episode/35gaKkxhNQ0WO9VyWU4RJ9): The Unenlightened Podcast, 2024. A two-hour conversation on the state of LLMs: the OpenAI, Google, Microsoft, and Apple AI keynotes, the hallucination problem, how AI affects society, and whether LLMs lead to general intelligence.
- [The Evolution of ONNX: Driving Interoperability in AI Through Open Standards](https://www.youtube.com/watch?v=N0wDg5oMHSA): AI.dev 2023 (Linux Foundation), San Jose, 2023. With Sai Kishan Pampana (Intel): how ONNX grew from an interchange format into the interoperability layer of the AI stack, why fragmented frameworks and tools need open standards, and where the ecosystem goes next.
- [The New ONNX Model Zoo](https://www.youtube.com/watch?v=5irQREPHm1Q): ONNX Community Meetup 2023, NVIDIA HQ, 2023. Presenting the relaunched ONNX Model Zoo at the ONNX Community Meetup: how TurnkeyML powers a reproducible, continuously validated model corpus for the ONNX ecosystem.
- [The Power of Graph Neural Networks](https://ramkrishna.dev/talks): GroqDay, 2023. Why graph neural networks stress conventional accelerator architectures, and how Groq's deterministic LPU architecture handles the irregular compute and memory patterns GNNs present. (Recording no longer public.)
- [The World of Ads and its evolving landscape](https://ramkrishna.dev/talks): Industry talk, 2023. How machine learning reshaped the advertising industry's infrastructure and economics, and where the landscape is heading. (Recording no longer public.)
- [The ONNX Ecosystem](https://ramkrishna.dev/talks): Groq developer session, 2022. A tour of the ONNX ecosystem - the interchange format, ONNX Runtime, ONNX Script, and the Model Zoo - and how they fit together for teams moving models across frameworks and hardware. (Recording no longer public.)

## Research

- [Patent: Hardware and software coordinated cost-aware low power state selection](https://patents.google.com/patent/US12141015B2): US 12,141,015 (granted), Intel Corporation, 2024.
- [Patent: Technology to augment thread scheduling with temporal characteristics](https://patents.google.com/patent/US10915356B2): US 10,915,356 (granted), Intel Corporation, 2021.
- [Exploring the Use of Dataflow Architectures for Graph Neural Network Workloads](https://link.springer.com/chapter/10.1007/978-3-031-40843-4_48): ISC High Performance 2023 Workshops (LNCS 13999), 2023. GNNs bottleneck on non-contiguous memory access patterns that GPUs handle poorly; this work evaluates how dataflow architectures from emerging AI accelerators alleviate those memory and compute bottlenecks for graph neural network workloads.
- [A Comprehensive Evaluation of Novel AI Accelerators for Deep Learning Workloads](https://ieeexplore.ieee.org/abstract/document/10024028): IEEE/ACM PMBS Workshop at SC22, 2022. Evaluates dataflow-based AI accelerators from SambaNova, Cerebras, Graphcore, and Groq against diverse workloads - deep learning primitives, benchmark models, and scientific machine learning applications - characterizing where novel architectures beat GPUs and why.
- [Platform-Aware Resource Allocation for Inference on Multi-Accelerator Edge Devices](https://www.emc2-ai.org/assets/docs/virtual-20/emc2-virtual20-paper-11.pdf): EMC² Workshop (Energy Efficient Machine Learning and Cognitive Computing), 2020. How to schedule inference workloads across the heterogeneous accelerators (CPU, GPU, VPU) on a single edge device: a platform-aware allocation approach that improves throughput and energy efficiency over naive placement.
- [Design and implementation of a semi-autonomous waste segregation robot](https://ieeexplore.ieee.org/abstract/document/7506679): IEEE SoutheastCon, 2016. A semi-autonomous robot that identifies and sorts recyclable waste - early work combining computer vision with embedded robotics.

## Coverage

- [One vLLM Stack, Two Architectures: Prototype on RDNA, Scale to an MI300X for the Price of Lunch](https://netstatz.com/vllm-rdna-cdna-bridge/) (Netstatz), 2026: “Ramakrishnan Sivakumar at AMD authored the original vLLM backend, and Jeremy Fowers led the refactor.”
- [Lemonade 11.5 Local AI Server Released With Completed Lemonade Router](https://www.phoronix.com/news/AMD-Lemonade-11.5) (Phoronix), 2026: “The most significant change with Lemonade 11.5 is the completion of the Lemonade Router that can be used for automatically routing queries to relevant models based on defined policies.”
- [I switched my local AI setup to AMD's Lemonade after Nvidia support landed, and solved my local AI portability problem](https://www.xda-developers.com/local-ai-amd-lemonade-nvidia-support-solved-local-portability-problem/) (XDA-Developers), 2026: “Use your stack on whichever hardware you happen to be using at the time.”
- [Stop Crashing and Start Cooking with vLLM on AMD and Lemonade Server](https://pub.towardsai.net/stop-crashing-and-start-cooking-with-vllm-on-amd-and-lemonade-server-bef66caf5db0) (Towards AI), 2026: “How I Fixed vLLM on Strix Halo and Got 3x Better Batch Throughput with Qwen3.5.”
- [AMD-backed Lemonade local AI server adds NVIDIA CUDA support](https://videocardz.com/newz/amd-backed-lemonade-local-ai-server-adds-nvidia-cuda-support) (VideoCardz), 2026: “This is an interesting change for a local AI project backed by AMD, since Lemonade was primarily built around Ryzen AI NPUs, Radeon GPUs and x86 CPUs.”
- [AMD's Lemonade SDK For Local AI Adds NVIDIA CUDA Support](https://www.phoronix.com/news/AMD-Lemonade-10.7-Released) (Phoronix), 2026: “Lemonade, the local AI server solution developed by AMD that is designed to work across their CPUs, GPUs, and NPUs, is out with a new version today that also adds NVIDIA CUDA support.”
- [AMD's Lemonade SDK For AI Promotes macOS To GA Status, ROCm 7.13 Integrated](https://www.phoronix.com/news/Lemonade-SDK-10.5-Released) (Phoronix), 2026: “The Lemonade SDK for ‘refreshingly fast local AI’ that is largely developed by AMD engineers as an open-source project continues advancing quite rapidly for serving optimized LLMs on GPUs and NPUs.”
- [First look: Lemonade serves up local AI with limitations](https://www.infoworld.com/article/4169474/first-look-lemonade-serves-up-local-ai-with-limitations.html) (InfoWorld), 2026: “Lemonade, created by AMD, is a server application plus GUI for running local AI models, similar to projects like LM Studio.”
- [Lemonade by AMD: a fast and open source local LLM server using GPU and NPU](https://news.ycombinator.com/item?id=47612724) (Hacker News front page), 2026: 572 points · 111 comments on Hacker News.
- [AMD demos Strix Halo running 8 AI models to argue over “Is a hot dog a sandwich?”](https://videocardz.com/newz/amd-puts-strix-halo-loaded-with-8-ai-models-in-a-bun-fight-is-a-hot-dog-a-sandwich) (VideoCardz), 2025: “AMD decided to find the best way to show off Strix Halo's capabilities. They chose to throw eight local LLMs into a group chat and ask the one question that may ruin family holiday spirit.”
- [Lemonade for GitHub Copilot](https://www.youtube.com/watch?v=HUwGxlH3yAg) (Local models in VS Code Copilot Chat), 2025
- [Hugging Face MCP Course: Lemonade Server unit](https://huggingface.co/learn/mcp-course/unit2/lemonade-server) (Hugging Face), 2025: “AMD Partnered with Hugging Face to provide a guide on how to accelerate our end-to-end Tiny Agents application using AMD Neural Processing Unit (NPU) and integrated GPU (iGPU).”
- [ONNX Model Zoo organization on Hugging Face](https://huggingface.co/onnxmodelzoo) (Hugging Face), 2023: 2,326 models published under the ONNX Model Zoo organization on Hugging Face - the corpus rebuilt on TurnkeyML and relaunched at ONNX Community Meetup 2023.

## Optional

- [Writing index](https://ramkrishna.dev/writing): all articles
- [Talks index](https://ramkrishna.dev/talks): recorded talks and panels
- [Research index](https://ramkrishna.dev/research): publications and patents
- [Highlights](https://ramkrishna.dev/highlights): curated cross-platform highlights