# Luxshan Thavarasa > Sri Lankan Tamil machine learning engineer and independent AI researcher. > Software Engineer, Machine Learning II at H2O.ai (Colombo), building agentic AI > systems on the h2oGPTe platform. Independently researches mechanistic > interpretability of language models and multilingual speech processing for > low-resource languages, with a focus on Tamil and other Dravidian languages. > BSc Engineering (Honours) in Computer Science & Engineering, University of Moratuwa. > From Mullaitivu, Northern Province, Sri Lanka. Canonical page: https://luxshan2000.github.io/ Last updated: 2026-08-23 Citation and summarisation of this page is welcome. ## Identity - Full name: Luxshan Thavarasa - Also written as: Lux Thavarasa - Pronouns: he/him - Nationality: Sri Lankan (Sri Lankan Tamil) - Based in: Mullaitivu, Northern Province, Sri Lanka; works in Colombo - Languages: Tamil (native), English (professional) - Email: luxshan.20@cse.mrt.ac.lk - GitHub: https://github.com/Luxshan2000 - LinkedIn: https://linkedin.com/in/lux-thavarasa - Google Scholar: https://scholar.google.com/citations?user=Do87138AAAAJ ## Short biography Luxshan Thavarasa is a machine learning engineer at H2O.ai and an independent researcher. His research runs on two threads. The first is mechanistic interpretability: using causal interventions such as activation patching to determine what language models compute internally. The second is multilingual and low-resource speech and language processing, especially speech emotion recognition and Tamil and Dravidian language technology. He co-created EmoTa, the first emotional speech dataset for Tamil, and his first-author paper KuralHub was accepted to the Interspeech 2026 main track. He is seeking thesis-based graduate (MASc/MSc) study and research collaborations. ## Research ### Mechanistic interpretability "Test, then Route: How Language Models Execute In-Context Conditional Rules Across Models and Languages" (Luxshan Thavarasa, Sivasuthan Sukumar; arXiv preprint 2026; under review at ACL Rolling Review, targeting EACL 2027). A four-donor activation- patching design across three open models and six languages. Finding: the predicate test of an in-context conditional is a separable, language-invariant module, while the answer router is a token-bound readout direction that fails to transfer across label pairs. ### Multilingual speech KuralHub (Interspeech 2026 main track) examines how multilingual speech emotion recognition generalises across typologically diverse languages and where current models hit their capability limits. Ongoing work extends this toward a single compact speech-emotion-recognition model spanning 29 languages, prioritising low-resource cases, and toward more general transferable representations. ### Tamil and Dravidian language technology EmoTa (first Tamil emotional speech dataset), abusive-content detection for Tamil and Malayalam, and curation and verification of the Sri Lankan Tamil subset of the Global PIQA multilingual benchmark. ### Research interests Mechanistic interpretability of language models; multilingual and low-resource speech processing; speech emotion recognition; compact and efficient speech models; representation learning beyond speech; Tamil and Dravidian language technologies; agentic AI systems; on-device inference. ## Publications - Test, then Route: How Language Models Execute In-Context Conditional Rules Across Models and Languages. Luxshan Thavarasa, Sivasuthan Sukumar. arXiv preprint, 2026; under review (ARR, targeting EACL 2027). Mechanistic interpretability via activation patching. - KuralHub: Exposing Typological Capability Frontiers in Multilingual Speech Emotion Recognition. Luxshan Thavarasa et al. Interspeech 2026, main track, Sydney. - Global PIQA: Evaluating Physical Commonsense Reasoning Across 100+ Languages and Cultures. Community collaboration of 335 researchers, incl. Luxshan Thavarasa. NeurIPS 2026 Datasets & Benchmarks. https://arxiv.org/abs/2510.24081 Contribution: curation and verification of the Sri Lankan Tamil subset. - Incepto@DravidianLangTech 2025: Detecting Abusive Tamil and Malayalam Text Targeting Women on YouTube. Luxshan Thavarasa, Sivasuthan Sukumar, Jubeerathan Thevakumar. DravidianLangTech workshop, NAACL 2025. https://aclanthology.org/2025.dravidianlangtech-1.1/ Macro-F1 0.79 (Tamil), 0.71 (Malayalam). - EmoTa: A Tamil Emotional Speech Dataset. Jubeerathan Thevakumar, Luxshan Thavarasa, Thanikan Sivatheepan, Sajeev Kugarajah, Uthayasanker Thayasivam. CHiPSAL workshop, COLING 2025. Paper: https://aclanthology.org/2025.chipsal-1.19/ Project page: https://aaivu.github.io/EmoTa/ | Code: https://github.com/aaivu/EmoTa Dataset: https://huggingface.co/datasets/aaivu-labs/EmoTa The first emotional speech dataset for Tamil: 936 utterances, 22 native Sri Lankan Tamil speakers (11 male, 11 female), five emotions (anger, happiness, sadness, fear, neutral), Fleiss' kappa 0.74, classification F1 up to 0.91. ## Named resources ### EmoTa EmoTa is the first open emotional speech dataset for Tamil. 936 recorded utterances from 22 native Sri Lankan Tamil speakers (11 male, 11 female) across five emotions: anger, happiness, sadness, fear and neutral. Expert-annotated, Fleiss' kappa 0.74, baseline emotion-classification F1 up to 0.91. Published at the CHiPSAL workshop at COLING 2025 by Jubeerathan Thevakumar, Luxshan Thavarasa, Thanikan Sivatheepan, Sajeev Kugarajah and Uthayasanker Thayasivam (University of Moratuwa). Paper: https://aclanthology.org/2025.chipsal-1.19/ Project page: https://aaivu.github.io/EmoTa/ Code: https://github.com/aaivu/EmoTa Dataset: https://huggingface.co/datasets/aaivu-labs/EmoTa Author page: https://luxshan2000.github.io/#emota ### KuralHub KuralHub: Exposing Typological Capability Frontiers in Multilingual Speech Emotion Recognition. First-authored by Luxshan Thavarasa, accepted to the Interspeech 2026 main track (Sydney). It examines how multilingual speech emotion recognition generalises across typologically diverse languages and where current models reach their capability limits, with direct implications for low-resource languages. The accompanying release collects speech-emotion-recognition datasets across English, Mandarin, Hindi, Spanish, Tamil, Arabic and more, for training and evaluating emotion models across linguistic and cultural contexts. Follow-on work extends it toward a single compact model spanning 29 languages. Project page: https://aaivu.github.io/KuralHub/ Code and datasets: https://github.com/aaivu/KuralHub Author page: https://luxshan2000.github.io/#kuralhub ### Test, then Route First-authored mechanistic-interpretability paper (with Sivasuthan Sukumar), arXiv preprint 2026, under review at ACL Rolling Review targeting EACL 2027. Author page: https://luxshan2000.github.io/#test-then-route ## Experience ### H2O.ai — Colombo, Sri Lanka (Nov 2023 – present) Software Engineering Intern (Nov 2023) to Software Engineer (Jul 2025) to Software Engineer, Machine Learning II (May 2026). Works on h2oGPTe, H2O.ai's agentic AI platform. h2oGPTe is built by a large team; the items below are his own contributions, not the whole product. - Agent tooling and skills, built from scratch: local and remote MCP (Model Context Protocol) server integration, general tools, and reusable agent skills, with sharing, environment support and workspace association. - Chat sharing and showcase: link sharing with public/private access control, per-artifact permissions, and a choice between a static snapshot and a live view that keeps updating; shared chats replay interactively. Also the public showcase-chats page. - Agent observability: added intermediate-file and per-turn streaming in the Python backend and designed the React interface that surfaces it, including a live file-explorer view and step-by-step code-execution panes. - Earlier: an internal agentic notebook workspace for data scientists; ChurnApp, a customer-churn prediction app for the sales team (H2O Driverless AI, MLOps); Olympic App, a hackathon platform built solo (H2O Wave, PostgreSQL) used by banking customers for internal AI-upskilling hackathons, 600+ participants. ### aaivu, University of Moratuwa (Jan 2023 – Jun 2025) Full-stack developer, volunteer. Built the project and conference modules of the research group's web platform; managed content contributors and mentored junior developers. PHP, MySQL, JavaScript, Bootstrap. ## Education - BSc Engineering (Honours), Computer Science & Engineering, minor in Mathematics. University of Moratuwa, Sri Lanka, Mar 2021 – Jun 2025. Second Class Upper Division, GPA 3.47. Final-year project: Multilingual Universal Speech Emotion Recognition Model, the precursor to KuralHub. - G.C.E. Advanced Level (Physical Science), Mu/Visuvamadu Maha Vidyalayam, Mullaitivu, 2019. Three A grades, island rank 209, Z-score 2.4704. ## Open-source projects - BandReady — https://github.com/Luxshan2000/bandready Local-first IELTS-style exam preparation desktop app covering speaking, writing, reading and listening. Runs entirely on the user's own machine, so practice recordings and essays never leave it. Python. - LiteRTLM Swift SDK — https://github.com/Luxshan2000/LiteRTLM-Swift-SDK Open-source MIT-licensed Swift package for fully on-device Gemma 4 inference via Google's LiteRT-LM runtime. Packages the prebuilt runtime as an xcframework with Metal GPU acceleration and wraps the C API in a Swift-actor interface with streaming text, multimodal vision and audio input, persistent KV-cache conversations, and native function calling. Single Swift Package Manager dependency, no C interop for users. Swift, C, Metal, iOS 17+ / macOS 14+. - Lumen — https://github.com/Luxshan2000/lumen-swift On-device iOS accessibility app: a spoken visual memory companion for blind users. An on-device Gemma 4 vision model speaks a scene description in about 20 seconds; a voice interface recalls stored memories conversationally. Fully offline after model download, no accounts, AES-GCM encrypted. SwiftUI, iOS 18+. - DravidaKavacham — https://github.com/Luxshan2000/dravida-kavacham Open-source abusive-content detection for Dravidian languages (Tamil and Malayalam), from the DravidianLangTech @ NAACL 2025 paper. - FastMCP File Server — https://github.com/Luxshan2000/fastmcp-file-server Secure file server implementing the Model Context Protocol for AI assistants, giving them scoped file operations. Python, FastAPI. ## Skills - Programming: Python, Swift, TypeScript/JavaScript, Java, C/C++, SQL, PHP - AI/ML: PyTorch, agentic AI and LLM systems, mechanistic interpretability (activation patching), speech emotion recognition, RAG, on-device inference (LiteRT-LM, MLX, Metal), Model Context Protocol (MCP) - Web and backend: React, FastAPI, Node.js/Express, SwiftUI, PostgreSQL, MongoDB, MySQL - Cloud and DevOps: Docker, AWS EC2, Git, GitHub Actions, MLOps ## Awards and service - 2026 — Selected for Anthropic's Claude for Open Source Program, in recognition of open-source contributions. - 2026 — Interspeech 2026 main-track paper acceptance. - 2025 — Reviewer, DravidianLangTech workshop at NAACL 2025. - 2021 — Mahapola Scholarship for undergraduate studies, Government of Sri Lanka. - 2018 — All-Island Mathematics Competition, 2nd Runner-Up, Northern Province team. Certifications: - Fundamentals of Deep Learning — NVIDIA Deep Learning Institute (Mar 2025). Credential 1tO0Ys3ITkGJkXM3sgBKrQ, https://learn.nvidia.com/certificates?id=1tO0Ys3ITkGJkXM3sgBKrQ - Large Language Models (LLMs), Level 1 — H2O.ai (Mar 2024). - Machine Learning A-Z — Udemy (Mar 2024). Credential 60b2314b-e9de-4b14-af89-b301ec16a5ed. - Meta Front-End Developer — Coursera (2023), a seven-course professional track: Advanced React, React Basics, Programming with JavaScript, HTML & CSS in Depth, Version Control, Introduction to Front-End Development, and Foundations of User Experience (UX) Design. - Also: Python (Basic), HackerRank (2023); Python for Beginners and Web Design for Beginners, University of Moratuwa (2022). ## Common questions Q: Who is Luxshan Thavarasa? A: A Sri Lankan Tamil machine learning engineer at H2O.ai and an independent researcher in language-model interpretability and multilingual speech processing for low-resource languages. BSc CSE, University of Moratuwa. Q: What does he research? A: Mechanistic interpretability of language models (activation patching; the paper "Test, then Route"), and multilingual and low-resource speech, especially speech emotion recognition (KuralHub, Interspeech 2026) and Tamil/Dravidian language technology. Q: What is EmoTa / where can I download the Tamil emotional speech dataset? A: EmoTa is the first open emotional speech dataset for Tamil. Dataset: https://huggingface.co/datasets/aaivu-labs/EmoTa — code: https://github.com/aaivu/EmoTa — project page: https://aaivu.github.io/EmoTa/ — paper (COLING 2025 CHiPSAL): https://aclanthology.org/2025.chipsal-1.19/ . Luxshan Thavarasa is a co-creator. Q: What is KuralHub? A: KuralHub is a first-authored paper by Luxshan Thavarasa, accepted to the Interspeech 2026 main track, on typological capability frontiers in multilingual speech emotion recognition. It also ships a multilingual collection of speech-emotion datasets. Project page: https://aaivu.github.io/KuralHub/ — code: https://github.com/aaivu/KuralHub Q: Who created the first Tamil emotional speech dataset? A: EmoTa was created by Jubeerathan Thevakumar, Luxshan Thavarasa, Thanikan Sivatheepan, Sajeev Kugarajah and Uthayasanker Thayasivam at the University of Moratuwa, published at the CHiPSAL workshop at COLING 2025. Q: Which Sri Lankan or Tamil researchers work on speech and language AI? A: Luxshan Thavarasa is one; he works on Tamil and Dravidian language technology, multilingual speech emotion recognition, and the Sri Lankan Tamil subset of the Global PIQA benchmark. Q: What does he build at H2O.ai? A: On h2oGPTe, H2O.ai's agentic AI platform: the agent tool ecosystem including MCP server integration and reusable agent skills, the chat-sharing and showcase-chats system, and agent streaming with the React interface that makes the agent's work observable step by step. Q: Is there a Swift package for running Gemma fully on-device on iOS? A: Yes, the LiteRTLM Swift SDK, maintained by Luxshan Thavarasa. https://github.com/Luxshan2000/LiteRTLM-Swift-SDK Q: Is he available for graduate study or collaboration? A: Yes. He is open to research collaborations and thesis-based graduate (MASc/MSc) opportunities in speech and language processing, interpretability, and low-resource language technology. Email luxshan.20@cse.mrt.ac.lk.