Agastya Seth is a Staff AI Engineer building agentic systems, LLM evals, and AI infrastructure. https://agastyaseth.tech Rebuilding motion-corrupted signals with a conditional VAE before predicting blood glucose - Finger-prick tests and implanted sensors are still how most people with diabetes track their blood glucose. Both are uncomfortable and expensive, which is why estimating glucose from a smartwatch’s optical sensor is such an appealing idea. This post covers work I did at ASU’s ACME Lab, in collaboration with a... Sat, 10 May 2025 00:00:00 -0700 https://agastyaseth.tech/2025-05-10-non-invasive-glucose-estimation-from-wrist-ppg/ https://agastyaseth.tech/2025-05-10-non-invasive-glucose-estimation-from-wrist-ppg/ What happened when we brought ReNO and RAFT from image diffusion to Pyramid-Flow - Reward-based alignment has done a lot for text-to-image (T2I) diffusion. Methods like ReNO and RAFT measurably improve prompt adherence and image quality. Text-to-video (T2V) is a much harder setting: every frame has to look good, and consecutive frames also have to agree with each other. This post is about a... Thu, 12 Dec 2024 00:00:00 -0800 https://agastyaseth.tech/2024-12-12-reward-alignment-for-text-to-video-models/ https://agastyaseth.tech/2024-12-12-reward-alignment-for-text-to-video-models/ Paper Review - RecNMP: Accelerating Personalized Recommendation with Near-Memory Processing Introduction &amp; Motivation In an age where digital content and products inundate every aspect of our lives, personalized recommendation systems have emerged as critical tools for filtering and directing relevant information to users. Powering a diverse array of internet services—from search engines and... Mon, 18 Mar 2024 00:00:00 -0700 https://agastyaseth.tech/2024-03-18-recnmp-accelerating-personalized-recommendation-with-near-memory-processing/ https://agastyaseth.tech/2024-03-18-recnmp-accelerating-personalized-recommendation-with-near-memory-processing/ Paper Review - Centaur: A Chiplet-based, Hybrid Sparse-Dense Accelerator for Personalized Recommendations Introduction &amp; Motivation In today’s digital landscape, the capability of platforms like Netflix, Amazon, and Facebook to offer personalized content has become a cornerstone of user engagement. These platforms deploy recommendation models at an immense scale within their data centers, navigating... Mon, 18 Mar 2024 00:00:00 -0700 https://agastyaseth.tech/2024-03-18-centaur-a-chiplet-based-hybrid-sparse-dense-accelerator-for-personalized-recommendations/ https://agastyaseth.tech/2024-03-18-centaur-a-chiplet-based-hybrid-sparse-dense-accelerator-for-personalized-recommendations/ Paper Review - HyPar: Towards Hybrid Parallelism for Deep Learning Accelerator Array I. Introduction The rapid ascent of artificial intelligence, particularly Deep Neural Networks (DNNs), has revolutionized various fields by offering remarkable accuracy, scalability, and adaptability. DNNs form the backbone of numerous cutting-edge applications, ranging from facial recognition systems and speech processing to... Mon, 26 Feb 2024 00:00:00 -0800 https://agastyaseth.tech/2024-02-26-hypar-towards-hybrid-parallelism-for-deep-learning-accelerator-array/ https://agastyaseth.tech/2024-02-26-hypar-towards-hybrid-parallelism-for-deep-learning-accelerator-array/ Paper Review - Distributed Deep Learning Using Synchronous Stochastic Gradient Descent Introduction In the landscape of deep learning, the capacity to efficiently scale up the training of large neural networks is a pivotal challenge. This paper tackles the optimization of synchronous Stochastic Gradient Descent (SGD), a cornerstone algorithm for distributed deep learning. The... Mon, 26 Feb 2024 00:00:00 -0800 https://agastyaseth.tech/2024-02-26-distributed-deep-learning-using-synchronous-stochastic-gradient-descent/ https://agastyaseth.tech/2024-02-26-distributed-deep-learning-using-synchronous-stochastic-gradient-descent/ Paper Review - SpAtten: Efficient Sparse Attention Architecture with Cascade Token and Head Pruning In revisiting the foundational concepts of Transformers and attention mechanisms, which we’ve discussed in previous articles, the introduction to the paper on SpAtten sheds new light on these critical elements in the landscape of natural language processing (NLP). This... Sat, 17 Feb 2024 00:00:00 -0800 https://agastyaseth.tech/2024-02-17-spatten-efficient-sparse-attention-architecture-with-cascade-token-and-head-pruning/ https://agastyaseth.tech/2024-02-17-spatten-efficient-sparse-attention-architecture-with-cascade-token-and-head-pruning/ Paper Review - Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity At the heart of today’s AI research, scale emerges as a monumental frontier. Observations have illuminated a simple yet profound truth: uncomplicated architectures, when backed by substantial computational resources, expansive datasets, and a vast number of parameters, consistently... Sat, 17 Feb 2024 00:00:00 -0800 https://agastyaseth.tech/2024-02-17-switch-transformers-scaling-to-trillion-parameter-models-with-simple-and-efficient-sparsity/ https://agastyaseth.tech/2024-02-17-switch-transformers-scaling-to-trillion-parameter-models-with-simple-and-efficient-sparsity/ Paper Review - Introduction Convolutional Neural Network (CNN) accelerators have emerged as a pivotal solution to the computational and energy demands of deep learning applications. These specialized hardware accelerators are designed to efficiently process the vast amount of data and complex operations involved in CNNs. Motivated by the need to enhance performance and... Mon, 05 Feb 2024 00:00:00 -0800 https://agastyaseth.tech/2024-02-05-eyeriss-an-energy-efficient-reconfigurable-accelerator-for-deep-convolutional-neural-networks/ https://agastyaseth.tech/2024-02-05-eyeriss-an-energy-efficient-reconfigurable-accelerator-for-deep-convolutional-neural-networks/ A Systematic Evaluation and Analysis of LLM Defense Strategies on Safety and Over-Defensiveness - The paper presents the Safety and Over-Defensiveness Evaluation (SODE) benchmark, which assesses Large Language Models’ (LLMs) defense strategies against safety concerns. It systematically evaluates and compares various strategies, revealing key findings like the trade-off between safety improvement and over-defensiveness in self-checking techniques, and the effectiveness of safety instructions in reducing... Mon, 15 Jan 2024 00:00:00 -0800 https://agastyaseth.tech/2024-01-15-the-art-of-defending-a-systematic-evaluation-and-analysis-of-llm-defense-strategies-on-safety-and-over-defensiveness/ https://agastyaseth.tech/2024-01-15-the-art-of-defending-a-systematic-evaluation-and-analysis-of-llm-defense-strategies-on-safety-and-over-defensiveness/ Govt. of India (MEITY) National Language Translation Mission - About Bhashini Bhashini aims to build a National Public Digital Platform for languages to develop services and products for citizens by leveraging the power of artificial intelligence and other emerging technologies. The project aims to create an ecosystem of academia research groups, industry collaborations, and crowd sourcing engines to facilitate... Wed, 12 Oct 2022 00:00:00 -0700 https://agastyaseth.tech/2022-10-12-contributing-to-the-bhashini-initiative/ https://agastyaseth.tech/2022-10-12-contributing-to-the-bhashini-initiative/ Automated Essay Scoring (AES) project - LiqVid<>SNU collaboration. - Experiments Python Notebooks The jupyter notebooks are described as follows: AES_MLP_BERT (9/10/2022) We compare results from 1D-CNN, 2D-CNN, and MLP training models with TF hub BERT APIs. Baseline-w-CSE (7/17/2022) Baseline training models for different experiments for the first stage (Semantic/Coherence/Prompt-relevance) with the spelling errors corrected from the essays to improve... Fri, 18 Feb 2022 00:00:00 -0800 https://agastyaseth.tech/2022-02-18-liqvid-snu-two-stage-automated-essay-scoring-aes/ https://agastyaseth.tech/2022-02-18-liqvid-snu-two-stage-automated-essay-scoring-aes/ My journey as a Data Scientist at one of the leading visual search startups - Wed, 19 Aug 2020 00:00:00 -0700 https://agastyaseth.tech/2020-08-19-visenze-project-report/ https://agastyaseth.tech/2020-08-19-visenze-project-report/ Melanoma Segmentation and Classification - Mon, 02 Dec 2019 00:00:00 -0800 https://agastyaseth.tech/2019-12-02-skinet-skin-cancer-detection-and-classification/ https://agastyaseth.tech/2019-12-02-skinet-skin-cancer-detection-and-classification/ Report for Major Project - I (Thesis) - GitHub https://github.com/agastyaseth/mimo-channel-estimation Sun, 01 Dec 2019 00:00:00 -0800 https://agastyaseth.tech/2019-12-01-massive-mimo-channel-estimation-using-deep-neural-networks/ https://agastyaseth.tech/2019-12-01-massive-mimo-channel-estimation-using-deep-neural-networks/ Report for my summer internship at Vi Dimensions 2019 - Sat, 03 Aug 2019 00:00:00 -0700 https://agastyaseth.tech/2019-08-03-vi-dimensions-project-report/ https://agastyaseth.tech/2019-08-03-vi-dimensions-project-report/ Winner - Hardware Edition - Tata Motors: Intelligent trip management for electric vehicles - SIH Hardware 2019 Problem Electric vehicles, especially pertinent to countries like India, suffer from a major disadvantage of not having charging stations as easily accessible as gas stations for regular cars. One of the biggest challenges in the adoption of electric... Fri, 12 Jul 2019 00:00:00 -0700 https://agastyaseth.tech/2019-07-12-smart-india-hackathon-winner/ https://agastyaseth.tech/2019-07-12-smart-india-hackathon-winner/ Embedded Course Project - Objective The objective of our project was inspired from the latest development in motorbike technology - self balancing. The idea was simple - make a prototype of a motorbike that would stand on its axis, and would counteract any swivel from that position. The working principle involves various sensors to... Thu, 23 May 2019 00:00:00 -0700 https://agastyaseth.tech/2019-05-23-self-balancing-bike/ https://agastyaseth.tech/2019-05-23-self-balancing-bike/ VLSI Course Project - Wed, 01 May 2019 00:00:00 -0700 https://agastyaseth.tech/2019-05-01-analog-vlsi-implementation-of-support-vector-machinelearning-and-classification/ https://agastyaseth.tech/2019-05-01-analog-vlsi-implementation-of-support-vector-machinelearning-and-classification/ IEEE 802.15.7 Visible Light Communication: Modulation Schemes and Dimming Support | Digital Communication Project - Tue, 30 Apr 2019 00:00:00 -0700 https://agastyaseth.tech/2019-04-30-lifi/ https://agastyaseth.tech/2019-04-30-lifi/