Summary
An ambitious Machine Learning Engineer with 7 years of experience and substantial success in designing and developing machine learning-based applications. Uses strong knowledge of the SDLC and Agile methodologies, stays abreast of emerging technologies, and delivers projects on time. Applies an analytical approach to solving problems and collaborating effectively in team settings. Aspiring Computer Scientist and researcher with a focus on Data Science, Machine Learning, and Deep Learning.
Overview
10
10
years of professional experience
Work History
Tech Lead
Arcadea Group
06.2023 - Current
- Machine Learning
- Utilize state-of-the-art LLMs to drive efficiency enhancements across Arcadea's and its portfolio companies
- Architect, implement, and optimize automation systems to streamline workflows and processes
- Conduct rigorous evaluations and employ monitoring techniques to ensure the optimal performance of machine learning models in real-world scenarios
- Actively work towards reducing biases in AI-driven outputs, ensuring the deliverables are fair, reliable, and unbiased
- Take proactive measures to avoid hallucinations in AI responses, ensuring the utmost accuracy and relevance in results
- Harness the capabilities of tools and platforms such as LangChain, PromptFlow, Azure, and OpenAI to craft and refine solutions tailored to Arcadea's objectives
- Collaborate cross-functionally with teams to understand needs, iterate on feedback, and deliver machine learning solutions.
01.2022 - Current
- The ability to predict the estimated arrival time (ETA) food delivery time is an important factor for customers, merchants, and the business as well
- An accurate prediction can improve customer satisfaction and help the supply side by dispatching the delivery partners at the right time.XGBoost regression-based model using Near-Real time features using spark streaming [Spark Streaming, Python, XGBoost, pandas, Jupyter, Hive, HBase, Apache Spark, Hadoop, Play Framework]
- Driver Supply and Customer Demand predictions, Predicting passenger demand can provide more insight to the drivers to maximize their profit, reduce the driver's vacant time and minimize passenger waiting time for accepting a trip
- Marketplace forecasting enables us to predict user supply and demand in a Spatio-temporal fine granular fashion to direct driver-partners to high-demand areas before they arise, thereby increasing their trip count and earnings
- Spark Streaming, Python, XGBoost, pandas, Jupyter, Hive, HBase, Apache Spark, Hadoop, Play Framework]
- Reinforcement Learning Approach to Algorithmic Trading (Masters Thesis Project) 2020-2021
- The goal of Algorithmic trading in financial markets is to find the alpha trading strategy because they have to adapt to continuously changing market conditions while beating competitor strategies
- Reinforcement learning algorithms have been developed to solve challenging tasks that have complicated problem spaces and imperfect information and a dynamically changing environment
- Formulated algorithmic trading strategy by Reinforcement learning using Distributed RL framework
- Hyperparameter tuning with distributed fault torrent cloud training processes to find optimal Models for achieving state-of-the-art results
- PyTorch, Ray, RLlib, SigOpt, Weights & Biases, Google Cloud Platform]
- Revenue Management System for Travel domain 2017 - 2021 https://www.codegen.co.uk/revenue-manager
- The problem of revenue management is a common one in many capacity-constrained industries
- This involves business decision-making to maximize one’s profitability in a market by allocating the right inventory to the right customers at the right price through the right channel
- Carried out data wrangling, Data Analysis, Feature engineering, time series segmentation, unconstrained demand prediction, price optimization, and capacity control for solving problems in the travel domain
- Designed and implemented a machine learning system that predicts hotel booking demand with over 90% accuracy
- Price prediction with Random forest, Adaboost, Wide and Deep neural networks
- Demand prediction - data gathering, demand prediction using random forest / AdaBoost, multivariate time series forecasting with RNN - DNN- Linear regression model, Training, evaluation models with dynamic model structures and hyperparameters, Bayesian hyper-parameter optimization, Research on seq2seq models for multiple time forecasting, hierarchical time series prediction, Recurrent neural networks (LSTM, GRU) with attention
- Building production-ready scalable machine learning solutions using Kubeflow and TensorFlow Extended (TFX)
- Utilized the TensorFlow Data API for faster input pipelines, efficient GPU memory management for model training, model evaluation with TensorFlow Model Analysis, and high-performance and flexible model serving with TensorFlow Serving
- Developed a framework in Python, TensorFlow, and TensorFlow serving for time series prediction based on recurrent neural networks with automatic hyperparameter tuning that achieved above 90% with exogenous features.Which used to do price optimisations in hotel tour operator Revenue management system
- Created Kibana dashboards using for tracking key metrics
- Analyzed and created dataset using booking search log dataset from Elasticsearch for booking prediction [Python, pandas, Jupyter, TensorFlow, Keras, Scikit-learn, Kubeflow, Apache Spark, Hadoop, Angular, Spring, Java ]
- Credit Card Center Cardholder Behavior Analysis 2018 - 2021
- Developed a statistical-based approach to track Cardholder Behaviors to enhance customer engagement
- Implemented customer segmentation in various aspects based on their attributes allowing management to take strategic actions to drive customer value
- Forecasted customer transactions and used the results for strategic planning.Utilized statistical models such as SARIMAX and deep learning-based time seq2seq models to predict future segment transactions
- Developed a customer churn prediction system to develop tailored retention measures for churners and non-churners and action suggestions based on their financial situation
- Developed as Time to event prediction to have a flexible definition of churn and rank risky customers using Recurrent Neural networks
- Python, TensorFlow, Keras, Tensorflow Serving, Scikit-learn, ts-learn, Fuzzy logic, Hadoop, Apache Sqoop, Dask, Apache Spark, IBM db2, Angular, Spring, Java]
- Automatic Facial and Number Plate Recognition System 2018 - 2021
- Face recognition technology is well-suited for use in environments like schools and offices, where individuals are typically the same people day in and day out
- In these settings, face recognition can be used to track office hours and activities, as well as to authenticate individuals based on their facial features
- This technology can serve as an alternative to fingerprint sensors, providing a convenient and efficient way to verify identities
- The system is safeguarded with real-time liveness detection to prevent spoof attacks
- Android application that can quickly be installed in any environment for real-time recognition using on-device inference
- The system is designed to scale with multiple deep learning models based on people's activity density, making it efficient and effective in a range of environments
- Developed & deployed Deep learning based vehicle number plate detection system that will identify vehicle number plates from CCTV camera feeds
- Reduced inference time in deep learning vision models by 80% and GPU memory footprint by 30% using quantization and DL inference frameworks to scale up the face recognition in multiple CCTV camera systems
- Python, FastAPI, Angular, Android, OpenCV, MXnet, Multi-Model Server, Tensorflow, Tensorflow serving, TensorRT, RabbitMQ,Socket.IO, Raspberry Pi, Graphite, Grafana]
- Chatbot Natural language understanding (NLU)
Intern Software Engineer
GTO-SI
01.2021 - Current
- AI Initiative
- Intelligent Chatbots
- Worked with GTO-SI Research and Development team to develop domain-specific, intelligent, automated, virtual support assistants
- Python, NLTK, Python, Flask, Java, JavaFX, NodeJS, SWAGGER, AngularJS, IBM BlueMix, MongoDB, WSO2 CEP, iOS]
- Handwritten Character Recognition
- Worked with a small research team to conduct research and develop a handwritten character recognition (HCR) module for extracting details from bank cheques
- Python, OpenCV, TensorFlow]
- Dashboard for automating calls using text to speech [AngularJS, NodeJS, SWAGGER, Twilio API]
- Dashboard for showing real-time data for vehicle status [Android, Maven, NodeJS, SWAGGER ]
- PROJECTS
- LLM-Driven Agent-Based Automation for Aviation Parts Sales, This project combines Large Language Models and a specialized multi‑agent framework to streamline quoting, vendor management, and customer qualification on a Salesforce‑backed aviation parts platform
- Each agent handles a specific role—coordinating seamlessly to reduce manual effort and boost data-driven decision-making
- Multi‑Agent Coordination: Specialized agents (quote analysis, pricing, vendor management, customer qualification) handle discrete tasks and collaborate in a shared workflow
- Salesforce Integration: Direct interaction with Salesforce objects ensures real‑time synchronization of leads, opportunities, and quotes
- Intelligent Data Processing: GPT-based models interpret large volumes of internal data (quote history, part availability) and external sources (market trends, LinkedIn searches)
- Automated Decision Support: Agents leverage natural language processing to score potential customers and generate recommended pricing models
- Web Scraping + LinkedIn Search: Aggregates real‑time data on company growth, market positioning, and social signals
- Company Evaluation: Automated checks on vendor performance and financial stability inform quoting and partnership discussions
- Automated Quote Generation: Rapidly compiles customer details, part availability, and pricing logic into comprehensive quotes
- Tailored Follow‑Up: A dedicated agent schedules reminders and personalizes communication based on lead or customer segmentation
- Sequential Task Orchestration: Agents operate under a ‘crew and tasks’ model, ensuring quote analysis, pricing suggestions, and final reporting occur in a logical order
- Real‑Time Reporting: Dashboards visualize pipeline performance, showing opportunity stages, forecasts, and revenue trends at a glance
- Agentic AI
- Agentic workflows
- Multi-agent system]
- Becosoft PIM – Product Content Generation with AI
- Becosoft PIM is an AI-powered platform that automates the creation of high-quality, SEO-enhanced product descriptions and engaging blog content, streamlining the product listing process and boosting marketing effectiveness
- Key Features:
- Product Description Generation:Automatically creates compelling, SEO-friendly descriptions based on provided product details
- Blog Post Generation:Generates engaging blog articles by accepting a topic, desired tone and length, and a list of product names with links
- The system intelligently inserts relevant product links—turning blog posts into effective marketing tools
- Blog Image Generation:Produces a curated list of realistic images to complement and enhance your blog content
- Becosoft PIM Streamlines Content Creation.Users input product details or blog parameters (topic, tone, length, product links) into the system
- Azure OpenAI then generates tailored content, while FastAPI handles API requests and workflow orchestration
- The system utilizes Cosmos DB to manage data storage and Azure Functions to power a serverless, scalable architecture, ensuring cost-effective performance
- By automating content creation tasks, Becosoft PIM enhances SEO and online visibility while reducing manual workload, allowing teams to focus on strategic growth
- Azure OpenAI, FastAPI, Cosmos DB, Azure Functions]
- AI-Powered Copilot API | Intelligent Knowledge Retrieval & Chatbot System, Designed and developed a flexible, AI-driven Copilot API that enables context-aware chatbot interactions by dynamically retrieving and summarizing information from various data sources (Salesforce Knowledge Bases, Google Docs, PDFs, and beyond)
- Originally conceived for aviation scenarios, this modular solution seamlessly adapts across industries, thanks to its RAG (retrieval-augmented generation) approach and secure, scalable architecture
- Key Features & Impact:
- Azure Cosmos DB for History Management: Stores conversation data, enabling smooth context switching and follow-up queries
- Azure App Service for Streaming (FastAPI): Ensures real-time, scalable interactions for 25+ concurrent users
- Dynamic Knowledge Retrieval: Connects to multiple repositories—Salesforce KB, Google Docs, PDFs—to provide accurate, context-relevant responses
- RAG-Powered Smart Responses: Injects the most relevant external documents into GPT-3.5 Turbo prompts, boosting the precision of chatbot answers
- Advanced Monitoring & Debugging: Integrates with LangFuse to capture detailed metrics, track errors, and enable iterative enhancements
- User Feedback Loop: Logs user interactions and applies insights to continuously refine prompt strategies and improve retrieval accuracy
- Scalable & Secure: Employs Bearer Token Authentication, optimized queries via Azure AI Search, and Azure OpenAI’s content filtering for compliance
- By leveraging Azure Cosmos DB for conversation history, Azure App Service for real-time API streaming, and a RAG approach for content retrieval, this Copilot API delivers context-aware, high-quality insights to end users—facilitating rapid deployment across diverse verticals and unlocking new possibilities for AI-driven knowledge retrieval and intelligent user interactions [Prompt Engineering
- Large Laguage mode
- Microsoft Azure
- Large language models
- Microsoft Azure Machine Learning]
- Smart Inbox & Email2Quote – AI-Powered Email Classification & Quote Extraction
- Developed a robust two-tier AI platform that streamlines aviation procurement by automating email classification and quote extraction with minimal human intervention
- Leveraging GPT-based models, Azure OpenAI, and LangChain, the solution employs an asynchronous, serverless architecture powered by Azure Functions and Azure Service Bus to handle high email volumes while adhering to OpenAI rate limits
- Key Features:
- Email Classification:
- Uses GPT models to categorize emails as Request, Response, or Junk and analyzes attachments (PDFs, images) to ensure proper routing
- Automated Quote Extraction:
- Utilizes LLM prompt engineering to extract part numbers, descriptions, conditions, quantities, and contact details
- Dynamically selects default or advanced models based on complexity
- Scalable & Resilient Processing:
- Employs Azure Service Bus for asynchronous request handling with robust error management and retry mechanisms
- Processes 100+ requests per minute via cost-effective Azure Functions
- Real-Time Monitoring:
- Integrates Power BI dashboards to track token usage, costs, extraction accuracy (90%+), and queue lengths for proactive capacity planning and cost optimization
- Reliable Data Management:
- Uses CosmosDB for persistent storage and Azure Blob for logging and debugging
- Impact:
- This end-to-end solution dramatically reduces manual data entry, accelerates quote turnaround, and enhances extraction accuracy while lowering operational costs
- It is ideally suited for modern generative AI deployments, empowering aviation procurement teams with faster, more reliable quote processing
- Python, FastAPI, LangChain, OpenAI, BeautifulSoup, Azure Functions, Azure Service Bus, CosmosDB, MongoDB, Power BI.]
- Food Recommendations
- In e-commerce and other digital domains, companies frequently want to offer personalized product recommendations to users
- Automatically recommending a product to a customer has become a fundamental part of many successful digital companies
- Designed & developed a Hybrid approach to generating recommendations by combining Content-based filtering & Collaborative filtering methods to serve hundreds of thousands of customers using spark and scala
- Improved customer clickthrough rate and conversion rate by 10X through continuous user feedback and click stream analytics by improving data quality and model performance tuning
- Developed word2vec based food tagging model for food item categorization and duplicate item identification [Spark MLlib, Python, pandas, Jupyter, Hive, HBase, Apache Spark, Hadoop, Play Framework]
- Estimated arrival time prediction for Ride-hailing and Food delivery
01.2017 - Current
- Apache Fineract platform Fineract is a mature platform with open APIs that provides a reliable, robust, and affordable core banking solution for financial institutions offering services to the world’s 2 billion underbanked and unbanked.Modifications for the client Deploying and maintaining website and databases on the AWS platform [Spring, Angular, MySQL, Gradle]
- Ontology modeling and querying via natural language (Final Year Research Project) 2017
- This project is an attempt to automate the knowledge structure building (automated ontology creation) while providing a natural language interface to query the said structured knowledge in the domain of disaster management.
Data Engineer
Arcee.AI
06.2024 - 10.2024
- LLM Training Data
- Designed and implemented scalable data pipelines using Metaflow and AWS Batch, enabling efficient processing and scaling of training data for Large Language Models (LLMs)
- Led efforts in NLP pre-processing, ensuring high-quality data preparation for training by normalizing, deduplicating, and validating datasets
- Collaborated with researchers and machine learning engineers to gather requirements and deliver datasets optimized for pre-training and fine-tuning LLMs
- Created and maintained a robust data pipeline for pre-training data creation, incorporating both public datasets and synthetically generated data
- Actively participated in customer calls to understand their requirements and provide technical support for data-related queries, enhancing client satisfaction
- Supported customer success teams by troubleshooting issues related to data integrity and pipeline workflows
- Stayed updated on best practices in MLOps and emerging tools for data engineering to optimize performance and scalability
- Key Skills and Tools:Python, SQL, Metaflow, AWS Batch, ETL Processes, NLP Preprocessing, Distributed Systems, Data Validation
Associate Tech Lead
PickMe (Digital Mobility Solutions Lanka (PVT) Ltd)
08.2021 - 07.2023
- Data Engineering and Data Science
- Working in the Data Science team to develop machine learning-based products in the Transportation & Mobility domain
- Designed and developed solutions to enhance the user experience, lower operational expenses, and increase sales metrics while adhering to data science best practices
- Influenced product roadmaps by providing insights through the data analysis and solutions to business problems in a dynamic environment
- Conducted Large scale data analysis, statistical inference, optimization, and linear & non-linear predictive modeling using data mining & machine learning
- Have established the data science practice and guided the data engineering & data science teams
- Designed and Developed resource-efficient Big data & Data science pipelines for ETLs, Analytics, and Machine learning use cases on the scale of terabytes of real-time and batch data using Scala and Spark
- Maintaining cloud Big data infrastructure using Hadoop, MapReduce, Pig, Hive, HBase, ZooKeeper, Kafka, Spark, Scala and Hue) to handle distributed big data workflows
- Handled Cloud migration from Azure HDinsight to Google Cloud Dataproc for Batch & streaming Data workflows
- Integrated Kibana and Elasticsearch for enhanced observability and logging, improving the data-driven insights
- Develop tools for data mining and visualization (Tableau Services , D3.js and Kibana)
Senior Software Engineer
CODEGEN INTERNATIONAL (PVT) LTD, LANKA
07.2017 - 07.2021
- Working in the Research and Development team to develop machine learning-based products in the travel and banking domain
- Conducted requirement engineering and solution design with direct client meetings for a reputed bank
- Containerized microservices, standardized Python packaging workflow, built, maintained, and monitored deployments, and GitLab CICD pipelines
- Built on-prem big data infrastructure using yarn, Hadoop, Spark, and Dask to handle distributed big data workflows
- Built and maintained complex ETL pipelines for data warehouses
- Configured multi-node GPU clusters to scale deep learning for model training & inference workloads
- Developed a framework in Python, TensorFlow, and TensorFlow serving for time series prediction based on recurrent neural networks with automatic hyperparameter tuning that achieved above 90% with exogenous features.Which used to do price optimisations in hotel tour operator Revenue management system
- Deployed Production-ready ML/DL models with optimized inference graphs for faster inference time and reduced GPU memory usage with quantization
- Reduced inference time in deep learning vision models by 80% and GPU memory footprint by 30% using quantization and DL inference frameworks to scale up face recognition in multiple CCTV camera systems
- Provided consultancy on technical standards, reproducible data science work, frameworks, tools, and coding standards to train and fine-tune deep learning/machine learning models in domains of time series prediction, vision, and natural language processing
- Developed scalable, fault-tolerant products using micro-services-oriented and containerized applications in Kubernetes Oracle (Managed Kubernetes service) and on-prem
- Designed and developed both relational database schemas (MySQL, MariaDB) and NoSQL database (Elastic Search)
- Robust application development using unit tests and SonarQube with maximum test coverage
- Member of company recruitment team hiring new employees, managing interns
01.2017 - 01.2018
- Https://www.lialive.ai/
- Developing a chatbot platform with back-office functionalities and pluggable AI bot features ‘Lia’
- Intent detection and semantic slot filling - word embeddings,seq2seq RNN models with attention
- Question answering models with BERT [Python, NLTK, TensorFlow, SyntaxNet, Protégé,Spacy, Rasa]
- Apache Fineract
VIRTUSA (PVT) LTD
01.2015 - 01.2016
Education
MASTER OF SCIENCE (MSC) - BIG DATA ANALYTICS
BACHELOR OF SCIENCE (HONS) - INFORMATION TECHNOLOGY
ADVANCED DIPLOMA - MANAGEMENT ACCOUNTING
GENERAL CERTIFICATE - ADVANCED LEVEL
Skills
Python
Java
Scala
R
NodeJS
PHP
JavaScript
MySQL
Oracle Database
IBM Db2
MS SQL Server
MongoDB
ElasticSearch
Hive
HBase
Gitlab-CI
Google Cloud Build
Git
SVN
GitHub
GitLab
Django
Flask
FastAPI
Spring
Hibernate
Angular 2
React
Node
Laravel
Maven
Gradle
Sonar
Logstash
FileBeats
Jira
Docker
Kubernetes
Oracle Cloud
AWS
GCP Dataproc
Azure HDInsight
Pandas
Numpy
Timeline
Data Engineer - Arcee.AI
06.2024 - 10.2024
Tech Lead - Arcadea Group
06.2023 - Current
-
01.2022 - Current
Associate Tech Lead - PickMe (Digital Mobility Solutions Lanka (PVT) Ltd)
08.2021 - 07.2023
Intern Software Engineer - GTO-SI
01.2021 - Current
Senior Software Engineer - CODEGEN INTERNATIONAL (PVT) LTD, LANKA
07.2017 - 07.2021
-
01.2017 - Current
-
01.2017 - 01.2018
- VIRTUSA (PVT) LTD
01.2015 - 01.2016
Robert Gordon University - MASTER OF SCIENCE (MSC), BIG DATA ANALYTICS
University of Moratuwa - BACHELOR OF SCIENCE (HONS), INFORMATION TECHNOLOGY
Chartered Institute of Management Accountants (CIMA) - ADVANCED DIPLOMA, MANAGEMENT ACCOUNTING
Royal College - GENERAL CERTIFICATE, ADVANCED LEVEL