Experience

Lead the AI-for-code sub-team in the WSO2 Integration team, building agent harnesses and small language models. Research focus: LLM failures in low-resource code generation and benchmarking LLM agents on enterprise integration.

• Designing multi-agent systems and developing/training custom models for agentic workflows.
• IntegrationBench: Leading a 100-task, behavior-graded benchmark that evaluates LLM agents by the runtime behavior of the integrations they build (targeting ICLR 2027).
• Domain-specific SLM: Developing a lightweight small language model for on-prem domain-specific code generation.
• Owning architecture, execution, and delivery of agentic developer workflows, and production readiness (security, scalability, reliability).
• Mentoring engineers and interns to ship high-quality AI features.

Led the effort to bring AI into WSO2 Micro Integrator.

• WSO2 Integrator Copilot: Owned from ideation to production. NL to Synapse DSL code generation with version-aware retrieval, XSD/Schematron validation, LSP diagnostics, and automated repair loops. 700+ active daily users.
• WSO2 Integrator Copilot Agent Mode: Cursor/Claude Code-style agentic integration engineer inside VS Code with 23 tools across 8 categories, multi-step planning, and autonomous task execution.
• Integrated domain-specific embedding models into WSO2 Integrator Copilot retrieval system for low-latency semantic code search.
• Optimized operational cost via advanced prompt caching (~90% cost reduction) and production hardening/observability hooks.
• Small Language Model initiative: Initiated domain-specialized SLM training for Ballerina to enable low-latency on-prem code generation.

Built open-source products for API Management and Integration.

• Delivered multiple end-to-end features for WSO2 API Manager and contributed 40+ product improvements.
• Distributed Transaction Counter: Scalable transaction counting for high-throughput API gateways handling 10,000+ TPS with sub-millisecond latency.
• GenAI / Multi-agent framework: Led a low-code framework supporting multiple LLM backends + vector stores, reducing delivery time and standardizing deployments.

• Developed architectures and POCs for a new API Marketplace intended to host and manage thousands of APIs.

• Developed end-to-end automated data analytics and Business Intelligence pipeline for the ZEPTO platform.
• Resulted in 40% improvement in decision-making speed for client companies.
• Built using Python, Scikit-learn, pandas, NumPy. Deployed with Docker and AWS.

Supervisors: Dr. Surangika Ranathunga (Massey University, NZ) and Dr. Nisansa de Silva (UOM)

• Leading Google-funded research on Automatic Post-Editing (APE) for low-resource languages (Sinhala, Tamil).
• Proposing controllable minimal-edit post-editing with hidden-state anchoring and dynamic edit-distance constrained decoding.
• Driving experimentation, evaluation, and research direction for practical APE workflows.

Supervisors: Prof. Sanath Jayasena, Dr. Miyuru Dayarathna

• Extended undergraduate thesis on privacy-preserving distributed graph ML.
• Developed novel common-attribute and common-entity aggregation mechanisms for heterogeneous multi-organization graph learning while maintaining privacy.
• Implemented prototype on JasmineGraph distributed graph database.

Supervisors: Prof. Sanath Jayasena, Dr. Miyuru Dayarathna

• Researched privacy-preserving distributed graph ML and federated GNNs for large-scale link prediction.
• Built memory-efficient distributed/federated GCN training enabling training on very large graphs on commodity hardware (JasmineGraph).
• Peer-reviewed publication at IEEE Big Data 2020.

• GPA: 3.87 - First Class Honours
• Dean’s List: 7 out of 8 semesters (Semesters 1, 3, 4, 5, 6, 7, 8)
• Thesis: Privacy-Preserving Machine Learning on Distributed Graphs (Grade: A+)
• University Award for Best R&D Project (highest marks achieved)

• First Runner-Up at ICASSP 2020 finals (team Icarus Inhibition).
• Achieved the highest score in the selection round for anomaly detection in autonomous drone systems using multimodal sensor data.

• Highest marks achieved in undergraduate thesis research on distributed graph ML.

• Awarded to the student who earned the highest marks for the semester 5 Software Engineering Project.
• Issued by Department of CSE, University of Moratuwa.

• DrivenData machine learning competition.

• Winning team (XGCoders) of nationwide inter-university data mining competition.

• 24-hour inter-university software development hackathon at University of Colombo.
• Team ODIGOS designed and developed a mobile/web app solving Sri Lanka’s tourism issues. First place out of 59 teams.