Ayush Agarwal

Machine Learning Engineer

Bengaluru, Karnataka, India6 yrs 5 mos experience

Most Likely To SwitchAI Enabled

Key Highlights

Expert in deploying ML models on edge devices.
Developed AI systems generating significant cost savings.
Pioneered LLM platforms for diverse applications.

Stackforce AI infers this person is a Machine Learning Engineer specializing in Fintech and AI applications.

Contact

Skills

Core Skills

Ai AutomationMachine Learning

Other Skills

AI AgentsMongoDBLanggraphAWS BedrockGCP Vertex AIAzure OpenAIAnthropicGraph Neural NetworksBlockchainNatural Language Processing (NLP)Graph Neural Networks (GNNs)NLPgRPCDatadogAutoformer

About

I'm Ayush, a Machine Learning Engineer specializing in ML system development and deployment. With nearly 4 years of experience in the field, I currently contribute to ML projects at Coinbase. I excel in deploying production-grade ML models on Edge Devices, including ARM-Based Machines, and have successfully developed speech models like TTS and ASR. Currently building ML applications in Web 3.0 space. Beyond my work, I actively share valuable insights and content on startups and AI on my LinkedIn profile. Join me in exploring the latest trends, best practices, and innovative applications in these domains. Skills: AI Agnets, MCP, A2A, PyTorch, TensorFlow, MLOps, NLP, Transformers, Apache Spark, Chatbot Development, GPT-3, Apache Airflow, ETL, Machine Learning, Deep Learning, Business Analysis, Business Models.

Experience

Coinbase

3 roles

Senior Machine Learning Engineer

Promoted

Aug 2025 – Present · 7 mos · Bengaluru, Karnataka, India · Remote

Developing a multi-agentic autonomous AI system to optimize compliance-critical case investigations and automate narrative generation from transactional data, ensuring regulatory compliance, reducing handling time by 70% and generating $12M in cost savings.

AI AutomationAI Agents

Machine Learning Engineer II

Apr 2023 – Jul 2025 · 2 yrs 3 mos · Bengaluru, Karnataka, India · Remote

LLM Platformization
Built the LLM platform at Coinbase to support various end-user and internal use cases, utilizing multi-cloud solutions (AWS Bedrock, GCP Vertex AI, Azure OpenAI, Anthropic).
Built a platform to enable developers to efficiently develop LLM powered applications in a self-serve manner with advanced retrieval and orchestration capabilities i.e RAGaaS, Knowledge Bases Creations, Multi-cloud Load Balancing for LLMs, etc.
AI Assistants / Agents for Compliance
Lead the overall development of building an AI Agent for efficient retrieval over 100+ complex PDFs for compliance agents.
Ideated, pitched and developed an PoC of Oncall Assistant for automatically performing root cause analysis of the alert.
Developed a preemptive filtering mechanism for IDV system using Multi-modal LLMs for real-time feedback for customers and cost-savings.
LLM Powered Use-cases
Detecting Fraudulent Transactions: Coinbase Scalable Blockchain Address Risk Scoring System
Developed a novel risk scoring system to generate risk score of blockchain addresses using Graph Neural Networks (GNNs) to predict potential fraudulent transactions and provide extended support for users.
Developed a novel and scalable approach to generate embedding of multi-billion nodes in blockchain transaction graph for both Ethereum and Bitcoin Blockchain using GNNs.
Filed a patent on the novel approach and published multiple artifacts like engineering blogs, research paper titled - "RiskSEA : A Scalable Graph Embedding for Detecting On-chain Fraudulent Activities on the Ethereum Blockchain".

MongoDBLanggraphMachine Learning

Machine Learning Engineer

Dec 2021 – Mar 2023 · 1 yr 3 mos · Bengaluru, Karnataka, India · Remote

Coinbase Chatbot (CBCB)
Developed Coinbase Chatbot (CBCB), an AI-powered conversational system designed to enhance customer support experience at Coinbase for all retail customers with an automation rate of 55%, resulting in substantial cost savings of $13 million dollars.
Developed a Zero-shot NLP Models with an accuracy of 91% in intent detection, enabling the chatbot to understand and respond to user queries effectively.
Designed and laid foundations to productionise the chatbot system by developing bi-directional streaming gRPC APIs, conversational engine for dialog flow, enabled analytics through instrumentation, monitoring and alerting via Datadog.
Fraud Detection of Blockchain Transactions
Developed a scalable blockchain analytics tool using GNNs to assess the risk of multi-billion blockchain addresses, achieving 78% recall in fraud detection significantly improving security.
Built ETL pipeline with Apache Airflow for data movement to/from Kafka, PostgreSQL, Data Lake and Snowflake.
Productionised the blockchain risk scoring feature generation / risk scoring pipelines using Spark Jobs in Databricks.
Developed auto-retrain ML pipelines to capture new fraudulent patterns with proper monitoring using Datadog and MLflow.
Developed real-time feature generation pipeline using Apache Flink.
Traffic Prediction Model
Developed a machine learning-based traffic prediction system using Autoformer, forecasting traffic 5 minutes in advance to optimize resource allocation. Achieved a 26% improvement in CPU utilization and reduced costs by 21% in instance minutes.
Integrated real-time data streams, deep learning models, and auto-retraining pipelines for accurate predictions.
Integrated solution with AWS Auto Scaling Groups (ASGs) and CloudWatch for proactive scaling and system stability.

Natural Language Processing (NLP)Graph Neural Networks (GNNs)Machine Learning

Jagriti yatra

Yatri

Nov 2023 – Nov 2023 · 0 mo · India · On-site

Jagriti Yatra is a 15-day long, national train journey that takes you 8000 kilometers across the length and breadth of India, to understand and build the India of smaller towns and villages through enterprise.

Agara

Machine Learning Scientist

Apr 2021 – Oct 2021 · 6 mos · Bengaluru, Karnataka, India · Remote

Developed a scalable and efficient Text-to-Speech (TTS) system for real-time conversations, using Transformers and Generative Adversarial Networks (GANs) trained from scratch for English language with an Mean Opinion Score (MOS) of 4.6 .
Led the development of end-to-end pipeline, including data collection from voice actors, model training, deployment using NVIDIA Triton, and evaluation with the help of human agents.

PyTorchStatisticsMachine Learning

Mihup

Machine Learning Scientist

Jun 2019 – Apr 2021 · 1 yr 10 mos · Kolkata Area, India

Text-to-Speech System
Designed and developed a real-time, offline multilingual Text-to-Speech system (Hindi, English, Bengali) with an MOS score of 4.05, optimized for edge devices like Raspberry Pi and ARM-based device.
Engineered a scalable and efficient streaming Text-to-Speech (TTS) engine, delivering high-quality voice synthesis optimised for real-time applications with an MOS score of 4.45 .
Automatic Speech Recognition
Developed, ported and optimised an ASR system for edge devices, utilising advanced model compression and optimisation techniques like pruning and quantization to improve efficiency and minimise resource footprint.
Platform and MLOps
Designed and developed a scalable, robust platform with bi-directional streaming APIs using gRPC to expose TTS, ASR, and NLU capabilities, powering multiple product offerings and enabling seamless integration across the organisation.
Technologies : C++, Python, ARM Neon Intrinsics, Neon Optimizations, Rust, Tract, Cargo cross-compilations, gRPC, Model Pruning, Quantization, Multi-threading, Thread-Safe Data Structures, Streaming, MongoDB

ArduinoPyTorchMachine Learning

Centre for product design and manufacturing, indian institute of science

Research Internship

May 2018 – Jun 2018 · 1 mo · Bengaluru Area, India

Designed and developed an algorithm to track the eye-gaze of a user , to perform certain gaze controlled on-screen operations on browsers in real-time.
Tested the system on students having certain speech and motor impairment on a sample quiz application.