Machine Learning Engineer
Email:
utkarsh.lal.official1@gmail.com

Utkarsh Lal

I currently work as a Machine Learning Engineer at Deloitte, building conversational agents and fine-tuning LLMs for RAG systems. I recently completed my MS in Business Analytics at UCLA, where I won first place in the Applied Analytics Program for building a Conversational Ads agent. Prior to Deloitte, I worked at Amazon, where I built an Autonomous Anomaly Diagnostics Engine using multi-step GraphRAG, processing data across 1000+ warehouses in production.
I have previously worked at American Express and Accenture, gaining exposure to everything from credit risk modeling and time-series forecasting to large-scale data engineering and production ML systems. I've also published research in Applied ML and Computational Neuroscience, which keeps me grounded in the fundamentals while working on applied problems.
What excites me most right now is making AI systems that actually work reliably in the real world. I'm drawn to the messy, practical challenges of turning research into products people can trust. My long-term goal is to lead teams building AI-native products that solve meaningful problems and push the boundaries of what's possible.

Research Interests

I'm currently conducting research in Computational Neuroscience. I'm interested in learning more about how AI tools can be leveraged to understand the neurological characteristics of debilitating disorders like Parkinson's and Alzheimer's. My recent research in this field utilizes Electroencephalography (EEG) and combines machine learning with statistical feature extraction measures such as fractal dimensions and entropies to capture the principal differentiating characteristics of such disorders. Furthermore, my recent research also delves into Sleep Stage Classification using Polysomnography, illustrating the significance of sophisticated feature extraction as a potentially more efficient avenue for sleep staging than complex deep neural networks.

Discover More

Professional Experience

KaggleX BIPOC Scholarship Program

August, 2023 - November 2023

StatusNeo

August, 2022 - Jan 2024

Data Science Consultant

American Express

March, 2022 - August, 2022

Analyst - Product Development

Accenture AI

August, 2021 - March, 2022

Data Engineering Analyst

Amazon

March, 2024 - Present

Data Engineer

June, 2025 - Sep 2025

Data Science Intern

Jan, 2025 - May, 2025

Data Science Intern

Deloitte

Jan, 2026 - May 2026

Machine Learning Engineer

Capital One

May, 2026 - Present

Senior Data Scientist

Research & Projects

Anomaly Diagnostics Graph Engine @ Amazon

• Led development of Anomaly Diagnostics Engine using graph-based fault detection and LLM-driven reasoning for 500+ robotics equipment in Amazon warehouses, reducing average hardware downtime by 50%
• Constructed knowledge graphs (HNSW) from unstructured troubleshooting manuals and 10 TB+ historical work orders, achieving 5–10s query latency across 1M+ nodes by building hybrid vector search query engine (FAISS), collaborating with Product and Software teams
• Developed and deployed multi-step graph RAG pipeline with Docker & AWS CDK in production, delivering context-aware troubleshooting and projecting a 50% MTTR improvement. (Neptune graphRAG toolkit, Langchain, LlamaIndex, Claude, Sagemaker)
• Built LLM-as-judge Entailment model to validate LLM responses, experimenting with Ragas for evaluation, reducing hallucinations

Research Paper: Fractal Dimensions and Machine Learning for the Detection of Parkinson’s Disease in resting-state EEG.

Doi: https://doi.org/10.21203/rs.3.rs-3270985/v1
Status: PUBLISHED in the Neural Computing and Applications Journal (IF: 6.0).
Supervision: Professor Arjun CV (PhD scholar in Technological University (TU) Dublin), and Dr. Luca Longo (Founder, AI and Cognitive Load Lab TU Dublin)

I conducted a comparative analysis of various Window segmentation, Machine Learning, and Fractal Dimensional techniques and proposed a novel model achieving over 97% accuracy in detecting Parkinson’s Disease from EEG. I also employed Explainable AI methods to enhance the interpretability of the model by visualizing feature importances yielded by the classification models using topographic plots of the brain. These plots accurately identified the motor cortex in the brain as having higher importance in differentiating between Healthy controls and Parkinson’s Disease patients. Furthermore, the proposed model illustrated robust performance in detecting Parkinson’s Disease in patients under varying levels of medication.

Research Paper: A comparative study on feature extraction techniques for the discrimination of frontotemporal dementia and Alzheimer’s disease with electroencephalography in resting-state adults

PUBLISHED in Brain Sciences

FTD is often misdiagnosed as AD. Therefore, there is a need for automated techniques to accurately differentiate between the two diseases. In this study, I built a model using SVD Entropy, sliding window segmentation, and Extreme Gradient Boosting for detecting and differentiating between Alzheimer’s and Frontotemporal Dementia with 90-95% average accuracy. I also employed XAI to identify relevant brain regions with higher degeneration by interpreting the results yielded by classification models. I also conducted a comparative performance analysis of several feature extraction measures and machine learning algorithms. This paper is currently going through the second round of review in IEEE Access. This paper also received funding from the Science Foundation Ireland for publication costs.

Sentiment Classifier and Web Crawler for product reviews

Client - Kimberly-Clark

Jan - Jul, 2021

Built a web application for scraping product reviews from Amazon.com and Shopee, paired with a sentiment analysis model.
Applied Lemmatization and Negation Handling in text pre-processing and utilized TF-IDF and n-grams for feature extraction.
Developed a POC for Topic Modelling using Latent Dirichlet Allocation to identify the most significant features in negative reviews and deployed the web app on Azure using Flask and Redis.

Audit Research Assistant @ Deloitte

• Enhanced Information Retrieval of Deloitte’s audit chat assistant, fine-tuning GPT-4o-mini using SFT on Azure OpenAI. Increased fact precision by 60% and deployed model on Azure AI foundry, collaborating with product & software teams
• Employed LangChain to build a ReAct Agent with session-based memory for an external client, serving over 200 beta testers
• Implemented prompting strategies like self-consistency and metric conditioning, increasing labelling confidence scores by 40%
• Integrated Semantic Deduplication by fine-tuning HuggingFace embeddings, reducing synthetic dataset redundancy by 20%

Research Paper: Ensemble Temporal feature extraction and Machine Learning for Classification of Sleep Stages from Telemetry PSG Data

Doi: https://doi.org/10.3390/brainsci13081201
Status: PUBLISHED in the Brain Sciences Journal (IF: 3.4)
Supervision: Professor Suhas Mathavu and Professor Anitha Hoblidar (Department of Electronics and Communications Engineering, Manipal Institute of Technology)

I implemented an ensemble feature extraction method using Power Spectral Density, Higuchi Fractal Dimension, Detrended Fluctuation Analysis, SVD Entropy, and Permutation Entropy, coupled with statistical measures, including standard deviation, kurtosis, skewness, and mean, to extract salient features of different sleep stages from Polysomnography data. Electromyography (EMG), Electrooculography (EOG), and Electroencephalography (EEG) biosignals were utilized. Comparative analysis of various Machine Learning models was conducted to determine the optimal pipeline for distinguishing between five different sleep stage configurations. The final pipeline achieved 90-97% accuracy across all configurations. This paper was published in the Brain Sciences journal.

Product Popularity Forecaster for Jio Mart

Client - Reliance JIO

Oct, 2022 - Present

Developed a SARIMAX model to forecast sales revenue and quantity of products available on an e-commerce platform called Jio Mart.
Decomposed time series to analyze trends and seasonality. Created exogenous variables to incorporate for increased user interaction during festive seasons.
Employed Custom Exponential Smoothing for handling seasonality and computed city-wise product popularities for JioMart.
Conducted hypothesis testing using the Augmented Dickey-Fuller test to validate the preprocessing performance.
Leveraged the pmdarima python package to build and tune the SARIMAX model.
Optimized the model and implemented mechanisms for logging and auditing.
Stack - python, pmdarima, statsmodels

End-to-End Data Architecture for Churn Analysis on Google Cloud

Built a churn prediction model using Call Detail Record (CDR) data of a telecom organization on Google Cloud's Vertex AI.
Created a streaming data pipeline handling 50,000+ rows per second of real-time data utilizing Spark Streaming.
Employed Recursive Feature Elimination and Ensembling with Random Forest and XGBoost using soft voting classifier.
Constructed batch Python data pipelines in Google Cloud Functions and DataProc. Orchestrated the pipelines using Apache Airflow and Cloud Composer. Performed extensive feature engineering using SQL in Google BigQuery.