Aug 2022 – Aug 2023 · 1 yr · Singapore · Hybrid

• Built Goal-Oriented Agentic Chatbot with Multi-Agent Reinforcement Learning: Architected a sophisticated conversational AI system using LangGraph and multi-agent frameworks comprising specialized agents for dialogue management, natural language understanding, and response generation. Implemented ReAct (Reasoning + Acting) patterns with tool-calling capabilities, enabling the system to autonomously interact with external APIs and knowledge bases. The Dialogue State Tracker (DST) leverages transformer-based policy networks and constitutional AI for context-aware responses, achieving 94% accuracy in intent recognition and task completion through RLHF (Reinforcement Learning from Human Feedback).
• Advanced Knowledge Extraction & RAG Systems: Developed GraphRAG architecture combining knowledge graphs with semantic retrieval for factoid question-answering. Implemented multi-hop reasoning using chain-of-thought prompting and tool-augmented generation, achieving 91% accuracy in complex reasoning tasks. The system integrates vector databases (Pinecone, Weaviate) with hybrid search combining dense and sparse retrieval methods.
• Predictive Analytics & Autonomous Forecasting Agents:
• Built Agentic Demand Forecasting Platform: Created autonomous ML agents that continuously optimize forecasting models using AutoML frameworks and neural architecture search. The system employs ensemble methods combining temporal graph networks, transformer-based time series models (TimeGPT, PatchTST), and multimodal fusion incorporating weather data, social signals, and economic indicators. Implemented real-time model adaptation with online learning and concept drift detection.
• Multi-Resolution Temporal Intelligence: Developed hierarchical forecasting agents capable of generating predictions from 15-minute intervals to yearly projections.

Nov 2020 – Aug 2022 · 1 yr 9 mos · Bangkok City, Thailand · Hybrid

• Meta Robyn & Advanced MMM Systems:
• Deployed Meta Robyn open-source MMM framework processing $2.3B annual media spend across 47 markets with 94.6% attribution accuracy and 99.7% model uptime
• Operationalized Bayesian hierarchical MMM using STAN and PyMC with Adstock transformations, saturation curves, and media interaction effects achieving 91% incremental revenue prediction accuracy
• Implemented advanced econometric methods (Ridge Regression, Sign-Restricted VAR, Penalized VAR, MLE) with Temporal workflow orchestration for real-time media optimization serving 340+ brands
• Built causal inference pipelines using LOESS smoothing, structural equation modeling, and difference-in-differences achieving 89% lift measurement precision across multi-touch attribution
• Architected Vector Error Correction Models (VECM) and Structural VECM with constitutional AI validation processing 15TB daily marketing data across 180 media channels
• Deployed Bayesian nonparametric models with Gaussian Process priors and Dirichlet Process mixtures for marketing mix decomposition achieving 97.3% forecasting accuracy
• Implemented advanced time series architectures combining Prophet, Neural Prophet, and TimeGPT with multi-modal fusion incorporating competitor intelligence, economic indicators, and consumer sentiment
• Built hierarchical Bayesian MMM with media saturation curves, adstock decay, and interaction effects using A2A protocol integration for cross-platform measurement
• Deployed gross margin optimization systems leveraging complex relational endogenous/exogenous variables achieving $127M incremental profit through dynamic pricing strategies
• Built autonomous marketing agents using MCP-enabled model routing with gradient boosting, bagging, and neural architecture search delivering 68% improvement in promotional ROI
• Implemented causal ML frameworks (DoWhy, EconML) with Temporal coordination for marketing experimentation and A/B testing

May 2018 – Dec 2019 · 1 yr 7 mos · Singapore · On-site

• Working as a lead solution architect for Walmart Labs (client) for building GT3 NexTech is the Data Science & Leveraged Engineering team under the Enterprise Business Services organizations. Enterprise areas of Walmart like Finance, Indirect Procurement, People Systems, and others are during massive digital transformation. Our team is building products for these areas from scratch, dealing with challenges from gathering user data, developing machine learning models to understand business processes better and making them more holistically data driven and intelligent. Automated Forecasting, Anomaly Detection, Object Detection and Segmentation, Chatbots, Optimizations are only some of the areas we are actively working on.
• Develop and automate reports and dashboard to provide insights on online user behavior, spanning across major digital advertising platforms as well as our website and app.
• Collaborated with the global data reporting team, product analytic team as well as tech teams around correct reporting of marketing data.
• Measured result. from A/B and multivariate testing and provide recommendations to improve campaigns
• Worked on Bid Landscape Forecasting Model the goal of NGD bid landscaped forecasting is to forecast the bid distribution for any advertising campaign in the NGD exchange system. I have used a divide-and-conquer approach to solve this problem in real-time. An input campaign is decomposed into samples according to the attribute values in it. targeting profile. Using sample history stored in a novel bid star nee, we forecast the sample level distribution by a non-linear regression model.
• Finally, sample-level estimates are aggregated using a mixture of log-normal models to generate bid distribution estimation for the ad campaign. The proposed approach offers both scalability and generality. (We evaluate our NGD bid landscape forecasting

Jul 2017 – May 2018 · 10 mos · Singapore · Hybrid

• developed lots of advanced techniques and products for IBM Watson AI. Our mission includes designing new deep learning techniques and novel efficient product features to improve IBM Watson services and accelerate IBM technique transformation. The goal is to facilitate our internal and external business partners. Another mission is to push the boundary of state-of-the-art AI techniques for IBM and the public deep learning community. Our research and products cover natural language understanding, computer vision, sensor signal understanding, medical and digital health, etc.
• Computer vision and Deep learning related projects.
• Fine-grained object classification
• Deep model compression and acceleration
• Mobile scalable deep learning model design.
• Mobile based deep model optimization with efficient power and latency.
• Improved the running speed of deep module on mobile by 9x.
• Deliver the scalable deep learning model with 6Mb (fine-grained object classification) with more than
• 95% accuracy (field test in the wild conducted by the QA group )
• Submitted three patents and eight papers in the past one year. These topics covered deep model compression and acceleration, object classification, multi-task learning, deep learning based mobile privacy and latency optimization, face image analysis, data mining, etc.

Jul 2015 – Jul 2017 · 2 yrs · Pune, Maharashtra, India · On-site

• The ICMOS project will implement cutting-edge computer vision technologies and automation systems across Seagate's chip manufacturing pipeline, from wafer inspection to final quality assurance. The system will leverage deep learning, real-time image processing, and predictive analytics to optimize production quality and efficiency.
• Business Objectives
• Primary Goals:
• Reduce chip defect rates by 35% through automated visual inspection
• Increase manufacturing throughput by 25%
• Minimize human error in quality control processes
• Achieve ROI of 280% within 24 months
• Establish Seagate as an industry leader in AI-driven manufacturing
• Key Performance Indicators:
• Defect detection accuracy: >99.5%
• False positive rate: <0.8%
• Production line uptime: >98.5%
• Quality control processing time: Reduced by 60%
• Cost per chip: Reduced by 18%
• Technical Architecture
• 1. Computer Vision System Components
• High-Resolution Imaging Infrastructure:
• 4K+ industrial cameras with specialized lighting systems
• Multi-spectral imaging capabilities (RGB, IR, UV)
• Microscopic inspection modules (up to 10,000x magnification)
• 3D surface profiling systems using structured light
• AI/ML Processing Engine:
• Custom-trained Convolutional Neural Networks (CNNs)
• Real-time object detection using YOLO v8 architecture
• Semantic segmentation for defect classification
• Ensemble learning models for improved accuracy
• Edge computing infrastructure for low-latency processing

Feb 2010 – Jul 2015 · 5 yrs 5 mos

• As a part of Next-Gen analytics team at the Global Decisions Management group I work on solving complex business problems across geographies and verticals using advanced analytics. The team is a group of internal consultants with expertise in Big Data & Modern Machine Learning.
• My projects include the following:
• Business: Citi Retail Services, North America
• Technology: R, Hive, Java,Python,NLTK,Deep Learning,RNN,CNN,Tensor Flow,Open CV
• • Designed an algorithm to recommend big-ticket items to customers of a large home improvement retailer
• • Generated 10X lift in dollar value capture compared to user-user collaborative filtering
• • Worked on terabytes of data to dig out spend patterns of customers
• Business: Credit cards, Singapore
• Technology: R, SAS, Java
• • Analysed credit card customers’ spend behaviour to predict their next visit at the partner merchants
• • Responsible for end to end project life-cycle spanning initialization, feasibility, analysis, modelling, execution, tracking and financial impact
• • The recommendation engine helped build a merchant partnership strategy based on the profitability predicted by customers’ spend behaviour
• Member of the Cash Application team, which is an integral part of the SAP Leonardo digital innovation ecosystem, where we try to automate enterprise financial accounting processes (accounts payable and accounts receivable) using machine learning.
• Invented an innovative method to solve a complex problem in accounts receivable (matching incoming bank statements with open receivables) using graph theory and deep learning with pending patent.

Feb 2010 – Jun 2015 · 5 yrs 4 mos

• Data Science and Machine Learning champion, and evangelist for the firm
• Implement rapid-time-to-market and applied innovation for lines of businesses and IT involving Life and Annuities, Retirement Planning Services, Group Protection products and services
• Leading and performing both strategic and hands on work on predicting strong drivers for successful sales and effective marketing strategies through machine and deep learning
• Create highly innovative predictive model architecture that differentiates us from our competitors
• Implemented new models to predict when Short Term Disability will turn into Long Term Disability claims
• Implemented new models to predict when Life Disability Policies will be terminated by our customers
• Implemented new models to predict how much Term Life quoted policies will produce in terms of profits or loss
• Perform in-depth and sophisticated multi-dimensional sentiment, product and sales analysis on the products and market we serve
• Implemented advanced Lifetime Value (Pareto/Beta-Geometric/Gamma-Gamma) and machine learning models to predict top advisors in our distribution network who can close sales with the highest probability matched against the opportunities in the market they serve
• Architecting the next generation chatbot technology for our customer service center using deep learning
• Working on cutting edge innovative models for estimating mortality rates
• Working on innovative features to drive our retirement planning tools and calculators
• Fast turnaround of machine learning prototypes using DataRobot, as appropriate
• Doing heavy research on how to develop insurance tontine and crowdsourcing products by blending AI with blockchain technology