Purdue University

University of Chicago

Eastern New Mexico University

Graduate Student, Purdue University

West Lafayette, Indiana, United States

Obtained MS, Mathematical Statistics: June 2021

Courses taken: Bayesian Data Analysis, Linear Models, Design of Experiments, Statistical Methodology, Mathematical Statistics, Computational Statistics, Stochastic Processes, Inference, Probability.

PhD Student

I lead the ML & Data Science efforts at Marin Software. I am a full-stack DS - I prototype models  in jupyter notebooks, work with PMs to gather product req, build and ship product in Scala/Pyspark.

Tools: Keras/Tensorflow, Python (sk-learn, pandas, numpy), R, Spark MLLIB, Scala, Scalatra, ScalaJS

Cross-Channel budget allocation among Search/Social/Display ad campaigns to maximize reach

Modeling and generation of optimal bids for a portfolio of search keywords to maximize revenue constrained by fixed budget using convex optimization techniques ( Lagrangian, Primal-Dual, Levenberg–Marquardt algorithm )

Multi-armed Bandits for Dynamic Ad Pricing.

Implemented Facebook Budget Allocation Algorithm to allocate a Campaign budget across multiple adgroups to increase conversions

Implemented custom Collaborative Filtering using Spark mllib ALS to prospect for LookAlike audiences given an audience segment, advertiser or product. Given domain impressions, clicks, conversions of user cookies from a mobile audience, we construct a 100m x 1m  user-product explicit recommendation matrix that indicates cookie propensity for domain. Via ALS, & similarity scores, we are able to obtain 5x lookalike audiences given a cookie audience sample : MLLIB

Tail term bidding via clustering using Decision Trees, Random Forests on Spark with word2vec - a novel approach wherein we use regression trees for clustering! Head terms ie. keywords in the small training set with known bids and nlp features, are trained using regression trees to cluster similar keywords. The rather large test set with tail terms ie. textual features & unknown bids, is then assigned bids of the training cluster : MLLIB

Marin Software

Principal Data Scientist/ Machine Learning Engineer

Founding member of the Data Science team. 
Mentor team members in Machine Learning, Spark, mllib, Scalding, Scala. 
1. Implemented Retention & Engagement Analysis in Scalding & Spark
2. Implemented Logistic Regression (SGD & BFGS) models in Spark MLLIB to predict Engagement for AnchorFree's Android app, as a function of 100,000 Android Apps. Built a training pipeline from scratch => data collection => labeling => successful prediction => move the needle on business.
3. Designed several novel data-infrastructure tools in Akka - A multi-actor sqoop to concurrently move large tables & access logs to hdfs, realtime dashboard using Akka actors.
4. Build an A/B Testing framework to score mobile apps using Bayes Factor, implemented front-end via Scalatra and Scala-JS

AnchorFree

Senior Data Scientist / Machine Learning Engineer

June 2013-Oct 2014 : Data Scientist on the Cloud Infrastructure Team (Observability, Mesos)

Tools used: Commons Apache Math, COLT, R, Scala, Scalding (Scala atop Cascading for Map Reduce jobs over Hadoop), Finagle, Mesos, with some mustache & scalaJS for front-end work.

Contributions: 
1. Capacity Utilization Modeling  & Visualization of all production jobs running on the Twitter Mesos cluster - Extracted p95 cpu & p95 memory utilization data for ~50,000 jobs running on Mesos per day, for over 200 days with minutely resolution. Cleanup, de-dup, and curve-fit data to identify most expensive Mesos jobs, their growth pattern. Predict cpu/memory usage in the forthcoming quarters with 95% confidence intervals, for forward-looking data-center forecasts. Provide models on granular  (Mesos per-job) and aggregated (Mesos per-role, per-datacenter) basis. Modeling - Linear Regression, Least Absolute Deviation ( LAD using Simplex ). ARIMA, Neural Nets. Tools used:  Apache Commons Math libraries in Scala & Scalding, Mesos, Finagle, scala-js for front-end charts. Implemented & optimized continuos data modeling (latest models built every hour) to minimize compute time given excessively large datasets.
2. Datacube aggregating 32 billion metrics per hour from 1800 Twitter services - provides unparalleled visibility into growth of telemetry stats at Observability - built with Scala, Scalding, Finagle, Mesos
3. A 10,000 feet overview of my job: <a href="https://blog.twitter.com/2013/observability-at-twitter">https://blog.twitter.com/2013/observability-at-twitter</a> 

Oct 2012 - June 2013 : Ad targeting on the Revenue Insight Team
1. Contributions to Scalding, Twitter's OSS Map-Reduce Scala framework
2. Geocoding via JTS Topology suite, Netacuity & Maxmind data

Twitter

Data Scientist II

Lead Scala Quant - 

CVAR Decomposition: Designed & implemented massively parallel Scala code to compute the Conditional Value at Risk using a Default/No-Default valuation model, on a commercial portfolio of over 400,000 loans. The computations involve high dimension matrix operations ( using BLAS via the Linear Algebra library COLT ),  and custom Monte Carlo simulations.
Coded a fast Quasi Monte Carlo engine in Scala to deliver low-discrepancy Halton sequences, which compares favorably with Mersenne Twister, WELL and Sobol MC engines. This Scala process processes a 400K monthly commercial portfolio under half an hour, replacing a single-threaded C++ legacy process which averages 6 hours!

CAQF: I implemented Commercial Asset Quality Forecasts, Expected Loss forecasting algorithms in Scala to forecast exposures on over a million facilities. This concurrent Scala implementation replaces and outperforms existing Matlab legacy code by an order of magnitude.

Krishnan Raman