Ramchalam Kinattinkara Ramakrishnan

Machine Learning Engineer | Generative AI Efficiency | Model Compression

Generative AI Efficiency (Qualcomm AI Research): My current work focuses on creating efficient solutions for Large and Small Language Models (LLMs/SLMs), multimodal foundation models, diffusion large language models (dLLMs) and other advanced LLM inference techniques like speculative decoding, accelerating on-device training.

Model Compression (Huawei Noah’s Ark Lab): At Noah’s Ark Lab, Montreal, my work entailed developing and implementing neural network model compression techniques to improve computational performance (quantization, pruning, NAS across conv and transformer based architectures).

Academic Research (McGill University): I obtained a thesis-based Master of Science in Machine Learning, where my research was advised by Prof. Mathieu Blanchette, focussing on applying Reinforcement Learning techniques in Bioinformatics.

Passion Project: My passion extends beyond inference to areas like LLM distributed training, reinforcement learning, kernel optimization and other low level op implementation.

Beyond my work, I spend my leisure time playing badminton, ping pong, and passionately following Formula-1 and Manchester United FC.

news

Sep 26, 2025	Paper “OmniDraft: A Cross-vocabulary, Online Adaptive Drafter for On-device Speculative Decoding “ accepted at NeurIPS 2025. 🎉✨ [link]
Sep 26, 2024	Paper “Forward-Forward Algorithm for On-Device Learning “ accepted at NeurIPS 2024. 🎉✨ [link]
Jul 16, 2024	Patent “Selective neural network pruning by masking filters using scaling factors “ has been granted. 🎉 [link]
Apr 26, 2022	Paper “An Empirical Study of Low Precision Quantization for TinyML “ accepted at tinyML Research Symposium 2022. 🎉✨ [link]
May 01, 2020	Join Qualcomm (Qualcomm AI Research) as a ML Research Engineer on the Embedded AI team.
Sep 05, 2019	Paper “Deep Demosaicing for Edge Implementation “ accepted at ICIAR 2019. 🎉✨ [link]
Oct 18, 2018	Join Noah’s Ark Lab (Huawei), Montreal as a ML Research Engineer on the Accelerated Neural Technology (ANT) team.
Sep 05, 2018	Thesis “Rlalign: a reinforcement learning approach for multiple sequence alignment “ accepted at BIBE 2028. 🎉✨ [link]
Sep 05, 2018	Paper “Detection of Errors in Multi-genome Alignments Using Machine Learning Approaches “ accepted at BIBE 2028. 🎉✨ [link]
Sep 01, 2018	Graduated from McGill University, Montreal, Canada, MSc in CS (Thesis)
Oct 18, 2013	Join Bharat Petroleum Corporation Limited (BPCL), India, Mumbai as a Software Engineer for the LPG team.
Sep 01, 2013	Graduated from Model Engineering College, Thrikkakara, India (CUSAT), B-Tech in CSE

selected publications

2025

NeurIPS

OmniDraft: A Cross-vocabulary, Online Adaptive Drafter for On-device Speculative Decoding

Ramchalam Kinattinkara Ramakrishnan, Zhaocong Yuan, Shaojie Zhuo, and 4 more authors

In Advances in Neural Information Processing Systems, 2025

Abs PDF

Speculative decoding generally dictates having a small, efficient draft model that is either pretrained or distilled offline to a particular target model series, for instance, Llama or Qwen models. However, within online deployment settings, there are two major challenges: 1) usage of a target model that is incompatible with the draft model; 2) expectation of latency improvements over usage and time. In this work, we propose OmniDraft, a unified framework that enables a single draft model to operate with any target model and adapt dynamically to user data. We introduce an online n-gram cache with hybrid distillation fine-tuning to address the cross-vocabulary mismatch across draft and target models; and further improve decoding speed by leveraging adaptive drafting techniques. OmniDraft is particularly suitable for on-device LLM applications where model cost, efficiency and user customization are the major points of contention. This further highlights the need to tackle the above challenges and motivates the \textit“one drafter for all” paradigm. We showcase the proficiency of the OmniDraft framework by performing online learning on math reasoning, coding and text generation tasks. Notably, OmniDraft enables a single Llama-68M model to pair with various target models including Vicuna-7B, Qwen2-7B and Llama3-8B models for speculative decoding; and additionally provides up to 1.5-2x speedup.

2024

NeurIPS

Stepping Forward on the Last Mile

Chen Feng, Shaojie Zhuo, Ramchalam Kinattinkara Ramakrishnan, and 3 more authors

In Advances in Neural Information Processing Systems, 2024

PDF