AI ML 

• 'Determining Acuity Rank of a Patient' using EHR data and LLMs - Patent Filed in 2/2025.
  Built as a python based implementation deployed on Azure Kubernetes as a Service , this PoC implementation analyzed data from EHRs and historical messages from a patient to determine their true clinical profile and computing an 'Acuity Rank' .  This work largely focussed on Natural Language Processing and the implementation was done in python and used Azure Open AI LLMs.  A patent filing was done in 2/2025 based on the work done here.
  
  

• Condition Chat at KP: PoC of a context aware conversational platform (LLM + RAG) for diabetic patients at KP.
  This limited scope pilot/PoC was built using python and Azure Open AI  LLMs.  It implemented Retrieval Augmented Generation technique to reduce hallucinations. The patient facing experience was a conversational bot. The objective was to present information to sufferers of diabetes to education them on their chronic condition, the do's and don'ts and how to manage their condition.
  
  

• Intelligent Messaging at KP: PoC for Intelligent Messaging platform to reduce physician burnout.
  Built as a python based implementation deployed on Azure Kubernetes as a Service , this implementation analyzed messages from patients to determine if they really deserved to go to physicians or could be routed to other nurse practioners or pharmacy teams. The work was largely focussed on Natural Language Processing and the implementation used  frameworks like spaCy as well as Azure Open AI LLMs to get to a fairly high degree of accuracy.
  
  

• Sig-Resolver - a tool to automatically decode SIG 

This work was done to help patients with automated notifications to take their medications on time. The presciptions written by providers usually follows the standard SIG format. This is an industry standard but allows for significant variability within.  This work included implementations in python with the resulting models deployed to Azure. It used several NLP frameworks such as frameworks like spaCy as well as Azure Open AI LLMs to get to a fairly high degree of accuracy even for some very complicated SIG situations. 

• Get Care Now at KP: Conversational AI based Symptom Triaging with an on-demand video visit < 5 minutes.
  This work was done with the help of an external vendor ada health and the Machine Learning engine from them was implemented within the KP.org web portal. Patients start their journey and answer a few questions about their symptoms in a chatbot and the deterministic outcome would then take the patient based on their condition and severity to a Virtual Video Visit or Next Day Appointment or some Home Remedies. 

• Red-flag Detector at KP : NLP based detection of ‘red flag’ conditions in conversations with patients at KP.

The patient's messages were classified using Natural Language Processing to determine if there was some sort of a red flag situation. The operational workflows were set up to alert the right teams.  

At Jebbit as the VP of Data Science I was leading all of the data engineering and analytics initiatives. While a big part of our work focussed on building foundational analytics and data platform work we also built some predictive algorithms using python and scikit-learn for usescases around customer churn and retention. The work included plenty of EDA and the implementations included logicistic regression , decision trees , support vector machines,  K-means clustering algorithms. 

At Fractal Analytics I was the VP of Engineering for the Customer Genomics, a Hyper-Personalization product which was built for the asset management industry/vertical. As part of this implementation we built several deep learning algorithms including the whole family of Next Best Action algorithms. These were developed on NVDIA GPUs using Tensorflow, Keras frameworks and this included some convolution neural network implementations. 

At Cintell.net, as the CTO I was leading all of the technology implementation and a big part of what we built used Natural Language Processing framework. We implemented modules that did content matching and contact matching as part of the 'Cintelligence' platform. These used frameworks such as nltk and models based on word embeddings such as Word2vec and other distance metrics such as Levenshtein distance and Jaro–Winkler distance and Sentiment analysis. We implemented all these to process billions of tweets in real-time using Apache Storm and Apache Kafka.  Using scikit-learn for some other Cintelligence usecases included EDA and the implementations included decision trees , and K-means clustering algorithms.