PhD Researcher · Melbourne, Australia
Nirodya
Pussadeniya
I build multi-agent reinforcement learning systems that coordinate under pressure — and that can explain themselves to the people depending on them.
My PhD builds hierarchical MARL for human–AI teaming in emergency response, at Deakin and Coventry. Before it, four years shipping production LLM pipelines and agent frameworks in industry — the research and the engineering keep each other honest.
- Doctorate
- Deakin & Coventry cotutelle, 2025–28
- Focus
- Hierarchical MARL & human–AI teaming
- Recognition
- 4 awards APICTA · NBQSA · Best Paper
- Industry
- 4+ years LLM & agent systems
Multi-agent reinforcement learning that people can actually work with.
Doctoral thesis
Adaptive Human–AI Collaboration for Emergency Response
Emergency environments change faster than human operators can replan. A wind shift can rotate an active fire front ninety degrees in minutes; ember attacks open new ignition points kilometres ahead of the front. Autonomous drone teams can absorb some of that load — but most multi-agent systems optimise mission metrics while ignoring the operator's cognitive state, and end up adding work rather than removing it.
So the question driving my PhD is this: how can an adaptive hierarchical multi-agent system dynamically allocate tasks, reason about human mental states, and balance individual against team-level objectives, in order to reduce operator cognitive overload in high-stakes response? Answering it means putting reinforcement learning, Theory of Mind modelling and human factors research into one framework, and then testing it with real operators in the loop.
I'm early in this — two papers published, the ASTRA framework implemented and benchmarked, and the simulation platform well underway. The work ahead is aimed at AAAI, AAMAS and ICML.
-
01
ASTRA — hierarchical multi-agent RL
An Asynchronous Skill Transition and Reallocation Architecture. A high-level coordinator assigns a shared team skill alongside per-agent skills while low-level policies execute primitives, with event-driven skill termination, dynamic reallocation, transition-based skill discovery and asynchronous credit assignment — so agents coordinate while running on independent temporal horizons. Implemented and evaluated on Google Research Football, Overcooked-AI and the StarCraft Multi-Agent Challenge.
-
02
Controlled selfish behaviour
Cooperative MARL usually penalises individual deviation as a coordination failure. But operational doctrine says local field intelligence sometimes justifies it. I'm learning a state-conditioned policy for when an agent should override team objectives to answer a locally critical event — and how that deviation gets communicated back and reconciled with the team.
-
03
Theory of Mind for adaptive autonomy
Estimating operator workload, trust and intent from multimodal behavioural signals — intervention frequency, decision latency, reliance patterns, and where feasible physiological measures — so the system can tune its own level of autonomy to the operator's current cognitive capacity instead of a level fixed at design time.
-
04
Language models as the reasoning layer
Knowledge graphs and LLMs sitting over agent state: structured knowledge to keep agent reasoning inspectable rather than merely scored, and natural language to make an agent's intent legible to the operator it is working with. This line runs back through my CARA and Affective-CARA work.
-
05
High-fidelity simulation testbed
A drone-based bushfire response environment built in Unreal Engine 5 — simulation architecture, operator and administrator interfaces, communication middleware and scenario management — giving reproducible human-in-the-loop experiments that measure NASA-TLX workload, trust, SAGAT situation awareness and mission outcomes side by side.
-
Affective‑CARA: A Knowledge Graph Driven Framework for Culturally Adaptive Emotional Intelligence in HCI
A knowledge-graph-driven agentic framework that combines cultural emotion knowledge, valence–arousal–dominance annotations and reinforcement learning to produce emotionally sensitive, culturally adaptive responses — cutting cultural representation bias by 61% and improving sentiment alignment over baseline models.
-
CARA: A Hybrid Framework Integrating Swarm AI Agents and Knowledge Graphs for Advanced LLM Reasoning
Presented at the 6th International Conference on Advancements in Computing. A framework for strengthening the structured reasoning of large language models through agentic workflows, reaching state-of-the-art results on reasoning tasks.
Four further papers are planned from the PhD — the ASTRA framework, controlled selfish behaviour, Theory of Mind based adaptive autonomy, and the integrated human-in-the-loop evaluation — targeting AAAI, AAMAS, ICML and JAIR. The current list is on Google Scholar.
Work that has been recognised.
-
Asia Pacific ICT Alliance Awards — First Runner-Up
Tertiary Student Projects: Solutions · Bandar Seri Begawan, Brunei Darussalam
Regional recognition across the Asia Pacific for FootEdge Pro, an AI-powered soccer performance analysis tool automating event detection, player tracking and advanced match analytics.
-
National ICT Awards (NBQSA) — Silver
Tertiary Category · Sri Lanka
National recognition for innovation and technical development in sports analytics, and the qualifying route to APICTA.
-
Best Paper Award — ICAC 2024
6th International Conference on Advancements in Computing
For CARA, on the contribution of swarm agentic workflows and knowledge graphs to large language model reasoning.
-
Best Employability Skills Achiever — 2nd Runner-Up
Faculty of Engineering, University of Sri Jayewardenepura
Recognising employability, leadership and technical ability across the graduating engineering cohort.
Earlier: finalist at TadHack Sri Lanka 2022 as team lead, top three at INTELLIHACK Master, Most Popular Idea at INTELLIHACK 2, and finalist at CS‑INSL 2022.
Research
-
Monash University
Department of Data Science & Artificial Intelligence · Research Assistant
Designing a causal trust model for multi-agent reinforcement learning environments, pairing a discriminative teacher–oracle trust signal with a learned concept bottleneck. Building a neuro-symbolic concept layer and a causal trust graph so estimates generalise past a single teacher's behaviour, plus a domain adapter that carries the model across environment layouts.
Python · PyTorch · Causal inference · Symbolic reasoning · Multi-agent RL
Industry
-
Synacal
Machine Learning Engineer · Sri Lanka
Architected and deployed RAG pipelines on LangChain, LlamaIndex, Pinecone and AWS, powering knowledge-management and chatbot ecosystems with 40% faster response accuracy. Engineered agent frameworks in AutoGen, CrewAI and LangGraph that removed over 60% of manual intervention, and built Markaive — a multi-agent platform that accelerated market research and content generation threefold.
LangChain · LlamaIndex · AutoGen · CrewAI · LangGraph · FastAPI · AWS · PostgreSQL
-
Velaris
AI Engineer · Sri Lanka
Built transcript-analysis and note-generation modules on transformer summarisation and topic modelling, and optimised Fact Factory — an event summarisation engine running over millions of customer interaction records — to sub-second latency. Integrated LLM pipelines across multi-tenant systems using asynchronous microservices.
PyTorch · BERT · LDA · FastAPI · AWS · MongoDB · Hugging Face
-
ASCII Corporations
Machine Learning Engineer, contract · Sri Lanka
Automated marketing report generation with AI agents, lifting campaign efficiency and ROI, and deployed NLP models for data-driven marketing strategy on LLaMA 3 and Groq.
AutoGen · CrewAI · LangGraph · LLaMA 3 · Groq
-
Kainovation Technologies
Associate Machine Learning Engineer · Sri Lanka
Engineered InsurePULSE, an AI system for insurers spanning backend ETLs, report generation and cloud deployment. Built Docufy, a document query tool for HR firms on LangChain and Elasticsearch, a sensitive-data filtering layer for enterprise chat systems, and ML models across finance and medical domains.
FastAPI · Azure · PostgreSQL · Elasticsearch · Kafka · React
-
EchonLabs
AI Engineering Intern · Sri Lanka
Built information extraction and verification models for national identity documents, contributed predictive models to AI‑CORE, and co-authored a research paper on AI-driven ID card verification.
Education
-
Deakin University & Coventry University
PhD, cotutelle — Information Technology · Melbourne & Coventry
-
University of Sri Jayewardenepura
BSc Engineering, Honours · Colombo
Domains
Multi-Agent Reinforcement Learning · Hierarchical RL · Human–AI Teaming · Theory of Mind Modelling · Multimodal Learning · Large Language Models · Knowledge Graphs · Explainable & Trust-Aware AI
Environments & simulation
Unreal Engine 5 · StarCraft Multi-Agent Challenge · Google Research Football · Overcooked-AI · PettingZoo · Gymnasium · RLlib · Ray
Tools
Python · PyTorch · TensorFlow · FastAPI · LangChain · LlamaIndex · AutoGen · CrewAI · LangGraph · Hugging Face · vLLM · DeepSpeed · LoRA / PEFT · Weights & Biases · MLflow · Pinecone · Qdrant · FAISS · AWS · Docker · Kubernetes · PostgreSQL · MongoDB · Redis
I speak often about agentic systems, reasoning, and what actually survives contact with production.
Talks
What Data Taught Us Keynote
CENSOC — University of Sri Jayewardenepura
Bridging Academic AI Research to Real-World Applications
University to Industry Orientation — Faculty of Engineering, USJ
Agentic Reasoning Using GraphRAG with Neo4j and Vertex AI
Build with AI — GDG Sri Lanka
Reasoning Capabilities of Agentic Systems
Global AI Bootcamp — Colombo
AI Agents as Neuro-Symbolic Systems
MLOps Community — Sri Lanka
Future is Agentic Keynote
Google I/O Extended — GDG Sri Lanka
Can LLMs Really Reason?
Global AI Bootcamp — Colombo
Community
Co-Chapter Lead
Omdena Sri Lanka Chapter — led AI-driven community projects
IT Manager
SEDS Sri Lanka — Aeronautical Division
Let’s talk.
I’m always glad to hear from people working on multi-agent RL, human–AI teaming, or anything nearby — collaborations, reading groups, benchmark comparisons, or a good argument.