Apr 2022 – Mar 2024 · 1 yr 11 mos · Gurugram, Haryana, India

• [Product Recommendation Engine (PRE)]: Introduced personalization at a buyer‑product level to improve the browsing experience for more than 50 million Tokopedia buyers. Executed two‑stage development process to ensure the system’s robustness and scalability: [Stage 1] Generate most relevant products (out of ∼600 million products) for buyers at buyer segment x category ‑ level (CatBoost Ranking Model using buyer segment x product past orders/clicks and raw product features) and [Stage 2] Rank the generated relevant product as per buyers preferences at that point in time (Two Tower Model using buyer past order/click history, demographics & products price, title, category features). In offline evaluation, proposed RecSys surpassed existing solution (Best Seller algorithm) by ∼5% lift in Personalization Score@5, and in A/B test, PRE boosted Homepage Inspirational Widget CTR (+15%) and CVR (+10%).
• [POC ‑ Next Category Prediction (NCP)]: Designed a recommendation engine that can remember sequential patterns of user’s actions to recommend more relevant and new categories. Implemented Next‑item Recommendation with Sequential Hypergraphs research paper using user’s past interactions (Click, Order, ATC, Wishlist, Discussion & Micro Actions) & recency as features, and last purchased category as target label. In offline evaluation achieved HR@1 ∼20% on out‑of‑time validation set which out‑performed baseline solution (user x category affinity‑based) by ∼45% lift.

Apr 2021 – Apr 2022 · 1 yr · Gurugram, Haryana, India

• Homepage Personalization:
• [Quick Cart Checkout]: Introduced a feature to simplify online food ordering with a single click checkout option on the homepage (readymade carts). Created user tags to identify whether a user has an affinity towards restaurant*dish or dish or restaurant using the user’s past order history. Based on these tags, the top 5 carts (restaurants + dishes) were identified for a user. Deployed a daily cron spark job on Airflow to update cart recommendations for incremental users.
• [People Like You ‑ User Personalized Dish Recommendations]: Implemented a collaborative filtering based algorithm (LightFM ‑ Implicit feedback, matrix factorization based model) for user personalized dish recommendations using user’s past ordered history and user*dish interaction, user average order value.
• Customer Analytics and Segmentation:
• [Customer Intent Classification]: Created customer personas (explorers, repeaters, low/medium intent user) of those users who opened the app using various app interaction signals which customer leave behind in terms of search and clicks activities. Applied clustering techniques to identify these customer personas using session length, total # clicks, and property interaction etc. features

Jun 2020 – Apr 2021 · 10 mos

• Led the team of 4 Junior Data Scientists in developing a robust information extraction and retrieval system to be used by various pharma clients

Jul 2018 – Jun 2020 · 1 yr 11 mos

• Text Analytics Products:
• Study Design Optimizer: Built a fully automated & end to end Natural Language Processing (NLP) pipeline that leverages state of the art models (BERT) & transfer learning to classify text and extract explicit/implicit information using syntactic and semantic relationships from protocol documents (PDFs) to make recommendations on optimizing clinical trial design. Reduced manual effort by 50-60% for protocol designing process and accelerated clinical programs through more efficient trial designs.
• Design Intelligence: Developed an end to end product that combines entity recognition & linking (BioBERT/BioFlair), and weak supervised learning with deep neural networks to classify unstructured biomedical text from open source data (ClinicalTrials.Gov, PubMed) to expedite the clinical program planning and trial design process. Reduced time, effort & cost for clinical trial designing (weeks to hours) and automated insights to better inform future clinical trial investment.
• Study Design Optimizer and Design Intelligence Products featured on Cision PR Newswire
• Predictive modeling and Explainable AI Solutions:
• Enrollment Rate Analysis: Analyzed the impact of study design attributes and KPIs on Enrollment Rate (Patient per Site per Month) using Generalized Linear Mixed-Effects Modeling (GLMM) to account for both fixed and random effects at protocol level
• Enrollment Rate Prediction: Built a stack of machine learning models to predict the Enrollment Rate using various study design attributes and pre-trained biomedical embeddings

May 2017 – Jul 2017 · 2 mos · Mumbai Area, India

• Developed a Machine Learning model to predict driver refusal after job is awarded in order to take further action
• Implemented LightGBM method (0.94 AUC) over imbalanced data, having 90% accuracy using booking details and driver’s past trend
• Designed a Restful web API using the Flask microframework in Python and deployed it using WSGI server (Apache)
• Company replaced existing model with proposed one in order to avoid refusal and driver profiling owing to its feasibility and accuracy