Apr 2022 – Present · 3 yrs 11 mos · Seattle, Washington, United States

• Built RAG based LLM application (from research to deployment) for market products achieving ROI of ~2.5M per year
• Led the research and development credit card portfolio risk models at scale (10K+ variables, >1 TB size). Built time-series models such as LSTM, Transformers on transaction histories: used advanced NLP/LLM models (LLaMa, Mistral, Falcon etc.), Topic Models (LDA),Tf-Idf, Word2Vec for embedding customer email interactions, financial statements etc. Achieved a lift of 2.5% over previous model which translated to cost savings of ~10M/yr
• Built predictive maintenance models using Terabytes of sensor data from various devices (Baseband, RRU, SFP RiLink etc.) in Radio Access Networks. Extensively built/fine-tuned LSTMs, Transformer based models/LLMs such as Sensor-Transformers, Time-LLMs. Deployed in Databricks with an estimated cost savings of ~5M/yr
• Developed anomaly detection algorithms on student loan letters using advanced Deep Learning and NLP techniques such as Deep-Autoencoders, Entity recognition using BERT based LLMs

Apr 2017 – Mar 2022 · 4 yrs 11 mos · Dallas-Fort Worth Metroplex · Hybrid

• Developed algorithms for portfolio optimization of assets based on Quantum-inspired Computing for a European Banking client of Fujitsu. Improved optimization time by 10-15% over standard algorithms
• Built a personalized recommender system for high-dimensional data for golf equipment. Reduced median FRL metrics of players by an average of 5%
• Built models for plant yield prediction using various sensor measurements such as CO2, humidity etc. Improved yield by 7% for tomato farms
• Built predictive and optimization solutions to problems such as last mile delivery of parcels in crowded cities, electric vehicle fleet routing optimization and design optimization of race-cars
• Built models for predicting insurance service calls and optimally deploying trucks to incident locations. Reduced cost of truck management by 15%
• Built models working with US government to predict PPE and Vaccine requirements at various counties in US for optimal resource distribution during the pandemic
• Leveraged advanced ML/DL/NLP (Scikit-learn, Keras, Tensorflow, PyTorch, NLTK, Transformers etc.) models and MIP optimization algorithms. Also extensively used PySpark, H2O, Dask, AutoML, Wandb, Metaflow, AWS Sagemaker, Kubeflow etc.

Sep 2015 – Apr 2017 · 1 yr 7 mos · New York City Metropolitan Area

• Led the Batch Behavioral Model project to predict failures of large number of batch jobs (~1 million) which have a complex dependency structure. Reduced cost of operations for failure handling by 30%
• Developed an M/M/c Queueing theory model to prioritize and optimally allocate alerts received in realtime
• Developed a comprehensive ensemble model utilizing a diverse array of techniques including NLP methods (on log data) such as Word2Vec, TF-IDF, Topic Models and a variety of ML algorithms such as PCA, Auto-encoders, Logistic Regression, SVM, and Deep Learning models like LSTM

Feb 2014 – Sep 2015 · 1 yr 7 mos · New York City Metropolitan Area

• Developed a Hidden Markov Model (HMM) based home/work dwell prediction system to tag users’ Home, Work, Retail, Recreation places etc. for smarter ad-serving. The model provided a lift of 10% for conversion rate
• Designed and implemented a conflation algorithm to dedupe ~1 Billion locations in PlaceIQ database. This further improved lift by 4%. Clustering algorithms such as K-Means, DBSCAN were implemented at large scale
• Built Place Visit Rate(PVR) product to measure the rate of users visiting PlaceIQ defined locations. Developed a Map- Reduce based algorithm that computes PVR on ~100 GB of data everyday. Developed an A/B testing methodology to measure campaign performance using a metric called PVR lift
• Extensively worked with Spark, Hadoop, Hive, Amazon EC2, MongoDB, Redis and Amazon S3

Jan 2012 – Jun 2012 · 5 mos · Greater Bengaluru Area

• Built a fraud-detection system to distinguish genuine apps from malicious ones at Amazon App Store using models such as SVM, Random Forest, Logistic Regression and Deep Neural Networks. Precision and Recall of 80% and 70% respectively were achieved on the malicious users entering the system
• Various data sources such as developer profile, graphic content, third-party tool usage, computational resource requirements etc. were used
• Worked with Amazon cloud technologies such as SQS (distributed queues), SWF (workflow management) and S3 for data storage and processing

Jun 2010 – Nov 2011 · 1 yr 5 mos · Greater Hyderabad Area

• Built a fraud-detection system to detect malicious users/bots. Classification algorithms such as SVM, Random Forest and Logistic Regression were used to achieve Precision and Recall of 75% and 50% respectively of the malicious users
• Integrated the system into the core authentication platform of Ivy Comptech to log genuine users in and also execute additional steps to handle low-confidence malicious users identified by the algorithm