Hi, I'm Mehul !

Data Scientist at Grab, Singapore

I'm work in the Generative and Agentic AI space tackling business challenges. As a product analyst, I bring strong product sense to developing impactful digital solutions.

Latest experiment

Exploring WebGPU to generate graphics in browsers
Profile picture for work section
Meili Mountains, Yunnan, China

Connect with me

LinkedIn Icon
Instagram Icon
Email Icon
Twitter/X Icon
YouTube Icon
TikTok Icon

GenAI Projects

Exploring the frontiers of generative AI to solve real-world challenges. This portfolio itself was vibecoded in Cursor. My work spans from experimental prototypes to production-ready interfaces for LLM integration.

My technical skills are complemented by my strong product analytics expertise, working at Grab, SEA's super App. I thrive at the intersection of data science and business strategy, transforming petabyte-scale datasets into actionable insights.

Automated Instrumentation

Problem Statement

The manual creation of instrumentation specifications was a significant bottleneck, demanding extensive cross-team effort and time. This inefficiency led to potential inconsistencies and hindered overall development velocity. There was a clear need for an automated solution to streamline spec generation and boost productivity.

Instrumentation Generation App Interface
Streamlit App

Solution

Spearheaded the development of a Generative AI-powered tool, starting with a successful Proof-of-Concept using Streamlit and OpenAI. Defined project scope, led the core workstream responsible for spec generation logic, and collaborated with analysts to build the Python-based workflow. Developed a fullstack application and created key interface components. Deployed the application with integrated logging, observability, and vector database support. Finally, drove user adoption through presentations, training sessions, and active feedback collection post-launch.

Technologies Used

Python OpenAI Docker Nginx FastAPI Streamlit

Chat Data Mining

Problem Statement

Extracting actionable insights from vast customer chat data volumes was challenging due to limitations in scalability and handling diverse languages. Existing methods struggled to consistently surface meaningful topics needed by business and product stakeholders. A more robust and scalable approach was required for effective chat data mining across different verticals.

Chat Data Mining Technical Architecture
Simplified Modeling Flow

Solution

Managed the chat data mining initiative, employing a hybrid approach combining scalable topic modeling (BERTopic) with advanced LLMs (Graph RAG, GPT-4o-mini) for enhanced topic representation. Addressed large data volumes and multi-language complexities, incorporating specific data pre-processing and NER strategies. Delivered a dashboard to visualize mined topics and insights, enabling deeper data exploration for stakeholders. Formulated strategies to further scale the solution across various languages and verticals, addressing identified gaps.

Technologies Used

Python Pyspark BERTTopic OpenAI PowerBI