May 2023 – Dec 2023 · 7 mos · Hybrid

• Built multiple end-to-end AI projects from scratch including ML, NLP and Gen AI technologies and saved more 1 million hours of manual work for medical coders in ~2 years at Episource.
• 🔷 UnifiedNER – NLP Solution to find Face2Face, Telehealth and Provider metadata in medical charts
• Implemented an end to end NER solution to extract entities which can tell us whether the patient and the provider met face to face(F2F entities) or had an online interaction (Telehealth entities) along with entities for provider metadata from medical charts/pdfs
• Led the end to end development right from getting the raw data from Snowflake, working with the annotators for creating the dataset and building multiple NER models to deploying the model in production.
• UnifiedNER was used to replace 4 production models as it had 20%-30% better weighted F1 score along with a low inference time which saved 520,000 hours of manual reading work for medical coders in order to find if an encounter is a F2F or Telehealth encounter.
• 🔷 M.E.A.T Validation using Gen AI – Using LLMs like ChatGPT to clean medicine-disease mapping
• Implemented an end to end solution using LLMs like ChatGPT and Llama 2 to understand which medicine-disease mapping is correct and can be used to validate if a patient is suffering from a disease or not and which mappings can be removed from the knowledgebase to reduce false positives.
• Used multiple prompting techniques likes Zero Shot Prompting, Few Shot Prompting, Role Based Prompting and Chain of Thought(COT) Prompting for validating the medicine-disease mapping.
• Reduced 6.97% False positives by removing incorrect medicine-disease mapping using the best prompt with the highest accuracy with the ChatGPT API(GPT 3.5 Turbo Model) and LangChain.

Nov 2021 – May 2023 · 1 yr 6 mos · Hybrid

• 🔷 Autocoding Engine – Machine Learning Solution to autocode medical charts and reduce false positives
• Built an Autocoding Engine which provides a probability score(0-1) for a disease term found in medical charts and that score will decide if that disease will be captured by a human medical coder or not.
• With this machine learning solution, we were able to reduce 48% false positives at 0.05 threshold with no negative impact on the entire NLP Engine.
• Saved 499,200 hours of manual medical coding work by eliminating 48% FPs with Autocoding Engine!
• 🔷 Clinical Suspecting using ML – Machine Learning Solution to suspect HCC Conditions for patients
• Implemented a machine learning solution to predict HCC Conditions using 1 million patient’s data for the client.
• Performed feature engineering using patient’s historical data like procedures, labs, medicines and diagnosis to generate ~32,000 features and the task was to predict HCC conditions for the patient.
• Clinical Suspecting Engine was used to replace a rule based system to predict HCC Conditions as it was beating the rule based system by a F1 score of 26% along with prediction explanations.

Aug 2020 – Nov 2021 · 1 yr 3 mos · Mumbai, Maharashtra, India · Hybrid

• Worked with multiple teams to build end-to-end DS solutions across different domains for clients.
• 🔷 Customer Reactivation Campaign– Multiple Machine Learning solutions to lower customer churn rate
• Worked on solving a business problem related to reactivation of customers for a leading global specialty retailer of health and nutrition related products to engage and convert one-time buyers into regular customers.
• Implemented product association with Market Basket Analysis to find the combination of products that are frequently brought together by customers using algorithms like Apriori and Association rules.
• Analyzed customer behavior across channels, product categories, discounts, loyalty programs, etc. and performed customer segmentation with the help of K-Means Clustering to create customer segments based on product affinity, channel preference among other factors.
• Increased customer reactivation by 17% and repeat customer basket value by 9% for the customized messaging group as compared to the general group.