Isuru Wijesiri

Colombo, Sri Lanka

I am an AI researcher and engineer at WSO2 and the University of Moratuwa. I work on natural language processing, agentic code generation, and graph machine learning, and I care most about systems that hold up in production and in low-resource settings.

At WSO2 (AI R&D), I lead the AI work for enterprise integration. The main project is WSO2 Integrator Copilot, a coding agent that turns natural language into Synapse DSL and Ballerina code, with version-aware retrieval and execution-guided verification. It is used by 700+ developers a day. I wrote a technical deep dive on how it works, and a series, Harness Engineering 101, on building agents from first principles. On the side, I build One Code, an open-source agent harness that runs the Claude Code workflow on any model or provider.

With the University of Moratuwa I lead Google-funded research on automatic post-editing for Sinhala and Tamil, two low-resource languages. That work produced a constrained-decoding diagnostic for post-editing (Findings of EMNLP 2026, first author) and the trilingual EnSiTa dataset and benchmark (preprint, under submission). I also lead IntegrationBench, which grades LLM agents by whether the integrations they build actually run.

Earlier, I built a memory-efficient federated GCN for distributed link prediction that trains on billion-edge graphs on commodity hardware (IEEE Big Data 2020), and a self-supervised anomaly detector for autonomous drones using video and IMU data (IEEE Signal Processing Cup 2020, runner-up).

Interests: NLP, Agentic AI, Retrieval-augmented generation, Code/DSL synthesis, Federated Learning, Privacy-preserving ML, Graph ML

news

Aug 31, 2026	Published Harness Engineering 101, a blog series on building LLM agents and harnesses from first principles.
Aug 19, 2026	Paper accepted to Findings of EMNLP 2026 (first author): “Confident but Wrong: A Constrained Decoding Diagnostic for Low-Resource Automatic Post-Editing”.
Jul 31, 2025	Started as Lead Research Collaborator at University of Moratuwa on Google-funded research on Automatic Post-Editing for low-resource languages (Sinhala, Tamil).
Feb 28, 2025	Presented a technical deep-dive on blockchain architecture, consensus algorithms, and distributed ledger mechanisms at WSO2 Technology Conference 2025. Watch the talk.
Jan 14, 2025	Released lightweight code retrieval models (22M/33M parameters) on Hugging Face: L6 and L12 , achieving 97% Recall@10 for domain-specific code search.

latest posts

Aug 17, 2026	Harness Engineering 101
Mar 18, 2026	Context Beats the Model: Building a Domain-Specific Coding Agent

selected publications

Confident but Wrong: A Constrained Decoding Diagnostic for Low-Resource Automatic Post-Editing

Isuru Wijesiri and others

In Findings of the Association for Computational Linguistics: EMNLP 2026, 2026

Abs HTML

A black-box, inference-time diagnostic for automatic post-editing (APE) that varies an edit-distance penalty between free editing and copying the machine-translation output. The resulting TER-vs-penalty curve exposes two systematic failure modes of low-resource fine-tuning, Binary Collapse and Confident Miscalibration, and prescribes a fix. The work releases the first English-Sinhala ( 66k) and a new English-Tamil ( 39k) APE datasets.
Memory-Efficient Graph Convolutional Network-Based Distributed Link Prediction

Isuru Wijesiri and others

In 2020 IEEE International Conference on Big Data (Big Data), 2020

Abs DOI HTML

Introduces a memory-efficient, federated training scheme that partitions large graphs across JasmineGraph and runs parallel GCN workers. This lets a single commodity server train link-prediction models on datasets as big as DBLP-V11 (10+ GB) that conventional GCNs cannot load, and cuts training time by 3-5x on Reddit and Twitter graphs compared with standard, in-memory methods.
Anomaly Detection using Deep Reconstruction and Forecasting for Autonomous Systems

Isuru Wijesiri and others

arXiv preprint arXiv:2006.14556, 2020

Abs HTML

Self-supervised anomaly-detection system for autonomous drones: forecasted next video frame with a conditional GAN and IMU vectors with dual LSTM (autoencoder + forecaster); achieved 94% accuracy / 0.95 F1 on video, 100% accuracy / 0.98 F1 on IMU, earning Runner-up, IEEE Signal Processing Cup 2020.