Sathvik Raju — Data Scientist

About me

IEEE & NeurIPS member

MS, Electrical & Computer Engineering — UIC

Avid reader (collects more than reads)

Travel · Writing · Painting

"A curious mind huddled with Tell Me Why? books — that's where data science found me."

I'm Sathvik, a Senior Data Scientist at Microsoft with a decade of experience turning messy real-world data into intelligent systems that actually ship. My path wound from Electronics & Communication in India to a Master's in ECE at University of Illinois at Chicago, where probability theory and neural networks clicked something into place.

Before Microsoft, I spent years at OnPoint Solutions building ML infrastructure from the ground up — writing algorithms from scratch, architecting ETL pipelines, and deploying models that reduced industrial operating costs by tens of millions of dollars. At Microsoft, I work on Windows 365 and Frontline Worker products, applying transformer models, statistical forecasting, and LLMs to make cloud computing faster and smarter for millions of users.

Active in IEEE and NeurIPS, my airport reads are recent research papers. I believe learning never stops — and that's still the most exciting part of the job.

Core toolkit

Python TensorFlow PyTorch Azure ML AWS SageMaker SQL / Kusto Apache Airflow Transformers Bayesian Networks Signal Processing GPT / LLMs A/B Testing

Career

Where I've built things

Microsoft Aug 2022 – Present

Senior Data Scientist

San Francisco, CA

W365 Hibernate — Cloud Cost Optimization

Developed transformer-based + statistical hybrid models with Microsoft Research Substrate team for Windows 365 hibernation prediction.
Resulted in 40% reduction in COGS per user and $1M in annual savings at 90% efficiency.
Streamlined MLOps pipeline using Azure ML, Synapse, and Kusto for modularized, scalable deployment.

↑ $1M saved · 40% COGS reduction

Frontline Prestart — Login Prediction

Led end-to-end ML model to predict user login times, enabling sub-60-second logins for ~60% of Frontline users.
Developed user segmentation framework and MLOps solution in Azure ML for fast iteration.

↑ 60% of Frontline users impacted

Copilot Scenarios — LLM Engineering

Fine-tuned LLMs and built Prompt Engineering workflows for incident ticket summarization and topic labeling.
Built validation pipelines for GPT-3.5 generated outputs to ensure reliability at scale.

OnPoint Solutions Aug 2020 – Aug 2022

Senior Data Scientist

San Francisco, CA

Abnormal Behavior Detection & Recommendation Engine

Full-stack lead on a 10+ burner heater anomaly detection framework — from data ingestion through model deployment and maintenance.
Models: Gaussian Mixture, Autoencoders, Copula-based outlier detection. Deployed via SageMaker & Apache Airflow.

↑ $50K saved per anomaly detected

Vegetation Failure Forecasting

Predictive analytics for external failures across multiple electrical plants — preventing wildfires and massive downtime.
Signal processing, Probabilistic Neural Networks, custom loss functions; data pipeline on AWS.

↑ $10M+ in operating cost savings since deployment

OnPoint Solutions Nov 2017 – Aug 2020

Data Scientist

New York City, NY

ML Product (Cortex) & Energy Grid Classification

Built KNN, Random Forests, Neural Networks, Bayesian Networks, LSTM from scratch for the Cortex platform.
1D-CNN + XGBoost fusion for type-of-failure detection on energy grids, reducing unit downtimes.
Real-time NOx forecasting (Solex burner) using SVR, XGBoost, MDP, MPC — saving $1M in year one.

↑ $1M first-year savings · Cortex product shipped

Privacera Aug – Nov 2017

Data Scientist

Fremont, CA

NLP — Document Classification & NER

LDA-based topic modeling for document classification for a cybersecurity client.
Custom Named Entity Recognition for sensitive entity detection across document sets.

Highlights

Work that shipped

Windows 365 Hibernate Prediction

Transformer + statistical hybrid models for cloud VM hibernation, developed in partnership with Microsoft Research. Deployed at scale across Azure infrastructure.

Governing equation

Transformers Azure ML Kusto MLOps

$1M annual savings · 40% COGS reduction

Industrial Anomaly Detection Engine

Full-stack anomaly detection framework for industrial burner systems, using Gaussian Mixture Models, Autoencoders, and Copula methods — now the highest-revenue product at OnPoint.

Governing equation

GMM Autoencoders SageMaker Airflow

$50K saved per anomaly · 6 months maintenance time saved

Vegetation Failure Forecasting

Predictive ML system for electrical plant external failure prevention. Probabilistic Neural Networks on raw signal data with custom loss functions — preventing wildfires since deployment.

Governing equation

PNN Signal Processing AWS EC2 Feature Engineering

$10M+ in operating costs saved

Copilot LLM Pipelines

Fine-tuned LLMs with prompt engineering for incident ticket summarization and topic labeling at Microsoft. Built validation layers for reliable GPT-3.5 outputs in production workflows.

Governing equation

GPT-3.5 Fine-tuning Prompt Engineering LLMs

Deployed across Microsoft Copilot workflows

Smart Combustion Control (Solex)

Real-time NOx forecasting and control for the first single-NOx burner. SVR + XGBoost with Markov Decision Process optimization, deployed via PLC edge computing.

Governing equation

SVR XGBoost MDP Edge Computing

$1M saved in year one of adoption

Cortex ML Platform

Built the core ML algorithm library for OnPoint's Cortex platform — KNN, Random Forests, Neural Networks, Bayesian Networks, LSTM, and Deep AutoEncoding GMMs, all from scratch in Python.

Governing equation

Python TensorFlow DAEGMM Bayesian

Shipped as commercial product used by non-DS teams

Beyond the work

What makes me human

✍️

Writing

I write on Medium — mostly about data science, ideas I find interesting, and the occasional reflection. It's how I process what I learn.

Read on Medium ↗

✈️

Travel

Exploring new places keeps my perspective fresh. I believe the best data scientists are the ones who stay curious about the world beyond their screen.

🎨

Painting

An unlikely creative outlet for someone who lives in numbers — but pattern recognition works just as well with colour as it does with data.

On curiosity & learning

"My airport reads are usually recent ML research papers. Being an active IEEE and NeurIPS member keeps me honest about what's actually moving the field forward."

I collect more books than I read (a known bug, not a feature). But that curiosity — the same one that had me buried in "Tell Me Why?" encyclopedias as a kid — is what makes the work feel less like work.

Making data
think for itself.

Core toolkit

Where I've built things

W365 Hibernate — Cloud Cost Optimization

Frontline Prestart — Login Prediction

Copilot Scenarios — LLM Engineering

Abnormal Behavior Detection & Recommendation Engine

Vegetation Failure Forecasting

ML Product (Cortex) & Energy Grid Classification

NLP — Document Classification & NER

Work that shipped

What makes me human

Writing

Travel

Painting

On curiosity & learning

Let's talk.

Making datathink for itself.

Core toolkit

Where I've built things

W365 Hibernate — Cloud Cost Optimization

Frontline Prestart — Login Prediction

Copilot Scenarios — LLM Engineering

Abnormal Behavior Detection & Recommendation Engine

Vegetation Failure Forecasting

ML Product (Cortex) & Energy Grid Classification

NLP — Document Classification & NER

Work that shipped

What makes me human

Writing

Travel

Painting

On curiosity & learning

Let's talk.

Making data
think for itself.