Portrait of Sachith Abeywickrama

Sachith Abeywickrama

Ph.D. Student · Nanyang Technological University & A*STAR, Singapore

I am a Ph.D. student in Electrical and Electronic Engineering at Nanyang Technological University, carrying out my research at A*STAR with the Singapore International Graduate Award (SINGA). I am also an Agentic AI Researcher Intern at Huawei Singapore's Search and Recommendation Laboratory.

My research is on efficient, trustworthy, and adaptive AI for sequential data: time-series foundation models, interpretable-by-design models, and continual adaptation under real-world distribution shift. Before my Ph.D., I built production NLP and LLM systems at Rootcode, and studied Industrial Statistics and Mathematical Finance at the University of Colombo.

  • Time-series foundation models
  • Interpretable AI
  • Continual learning
  • Efficient sequence modeling
  • Agentic AI

News

Publications

Hover over a thumbnail to see the method figure. Click a title for the project page.

Conference papers

  1. NeurIPS 2026 ConceptTime: shape-based versus predictive-geometry concept spaces

    Predictive Surprise as Self-Grounding Concept Bottleneck for Interpretable Time Series

    Sachith Abeywickrama, Emadeldeen Eldele, Min Wu, Xiaoli Li, Chau Yuen

    NeurIPS 2026 · Main Conference, Sydney

    One frozen forecaster's predictive surprise segments the series, grounds concepts without annotations, and gives a built-in fidelity score for abstention. 97.91% on Epilepsy with 1% of labels.

    @inproceedings{abeywickrama2026concepttime,
      title     = {Predictive Surprise as Self-Grounding Concept Bottleneck
                   for Interpretable Time Series},
      author    = {Abeywickrama, Sachith and Eldele, Emadeldeen and Wu, Min
                   and Li, Xiaoli and Yuen, Chau},
      booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
      year      = {2026}
    }
  2. NeurIPS 2026 EntroPE: patch boundaries placed at entropy peaks

    Entropy Guided Dynamic Patch Segmentation for Time Series Transformers

    Sachith Abeywickrama, Emadeldeen Eldele, Min Wu, Xiaoli Li, Chau Yuen

    NeurIPS 2026 · Main Conference, Sydney

    Patch boundaries go where a tiny frozen model's next-step entropy is high, so transformers see temporally coherent patches. 3.2% lower MSE than PatchTST with 90.9% fewer parameters.

    @inproceedings{abeywickrama2026entrope,
      title     = {Entropy Guided Dynamic Patch Segmentation
                   for Time Series Transformers},
      author    = {Abeywickrama, Sachith and Eldele, Emadeldeen and Wu, Min
                   and Li, Xiaoli and Yuen, Chau},
      booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
      year      = {2026}
    }

Preprints

  1. Illustration: a convolution kernel over one agent's feature stack, then a graph convolution across collaborating agents

    From Recursion to Convolution: Parallel Self-Improvement for Multi-Agent Systems

    openJiuwen Team, Huawei · Core contributors: Zheng Wang, Yuang Liu, Sachith Abeywickrama, Yangkai Ding, Tao Yu, Xinyu Zhang, Hongbo Wang, Qi Ye, Xuefeng Jin, Zhangchun Zhao

    Preprint, arXiv:2611.20891, 2026

    ConvRSI swaps sequential recursion for intra-agent convolution over features and inter-agent graph convolution over the collaboration topology, in latent space. Versus RecursiveMAS on MATH-500: 5.04% better, 17.27% cheaper, 45.11% lower latency.

  2. Illustration: shifted embeddings transported back into local charts of a frozen embedding space

    PACT: Recurrent Domain Incremental Adaptation of Time-Series Foundation Models

    Sachith Abeywickrama, Peiliang Gong, Emadeldeen Eldele, Min Wu, Chau Yuen

    Preprint, under review, 2026

    Chart-specific residual transports and an identity-preserving gate let a frozen TSFM adapt to recurring operating conditions from 5% labels, with no condition identity at test time. Above 0.95 macro-F1 on a four-condition bearing-fault stream, with forgetting below 0.05 on all five benchmarks.

  3. Illustration: raw bytes from several protocol messages with inferred field boundaries

    NeurInferno: Field Inference for Unknown Binary Protocols Without Priors or Binaries

    Sachith Abeywickrama, Min Wu, Xiaoli Li, Chau Yuen

    Preprint, 2026

    A trace-only neural model recovers byte-level field boundaries in protocols it has never seen: 0.807 average boundary F1 across 12 held-out protocols, lifting recall from 0.538 to 0.815.

Experience

  1. Oct 2026 – Present

    Agentic AI Researcher Intern

    Huawei Singapore · Search and Recommendation Laboratory

    Research on agentic AI for search and recommendation systems.

  2. Aug 2024 – Present

    Ph.D. Researcher / Research Scholar

    Institute for Infocomm Research, A*STAR · Singapore

    Efficient, trustworthy, and continual-learning methods for foundation models under distribution shift and limited labels, with multi-GPU training on SLURM for time-series foundation models and open-source LLMs.

  3. May 2023 – Jun 2024

    Associate Machine Learning Engineer

    Rootcode · Sri Lanka

    Document QA, semantic search, NER, and open-source LLM applications for enterprise clients, shipped with FastAPI, Docker, evaluation workflows, and cloud infrastructure.

  4. Nov 2022 – Apr 2023

    Artificial Intelligence Engineer Intern

    Rootcode · Sri Lanka

    Conversational AI and semantic-search systems for e-commerce discovery and support.

  5. Jun 2022 – Sep 2022

    Research Intern

    Nanyang Technological University · Remote

    Data-efficient polymer property prediction with topological molecular features and persistent spectral fingerprints.

Education

  1. Aug 2024 – Present

    Ph.D., Electrical and Electronic Engineering

    Nanyang Technological University, Singapore

    Research with A*STAR on efficient, interpretable, and continually adaptive AI for non-stationary sequential data.

  2. Jan 2019 – Apr 2023

    B.Sc. (Hons), Industrial Statistics and Mathematical Finance

    University of Colombo, Sri Lanka

Honors and awards

  1. Jul 2026

    Second Runner-up, China-ASEAN Student Innovation & Entrepreneurship Challenge

    Guiyang, China

    For “Precision Drive Coming,” an autonomous humanoid robotics joint solution representing NTU. Selected among 32 finalists from 1,007 entries across 149 universities.

    Sachith Abeywickrama on stage as second runner-up at the China-ASEAN Student Innovation and Entrepreneurship Challenge
    On stage in Guiyang
  2. 2024

    Singapore International Graduate Award (SINGA)

    A*STAR

    Competitive Ph.D. scholarship supporting doctoral study at NTU and research at A*STAR.

Technical skills

Machine learning
PyTorch, TensorFlow, scikit-learn, Transformers, foundation models, LLMs, self-supervised learning, continual learning, generative AI
Trustworthy & efficient AI
Interpretable AI, out-of-distribution detection, few-shot learning, model adaptation, efficient sequence modeling, uncertainty estimation
Applications
Time-series modeling, NLP, semantic search, information retrieval, document QA, named-entity recognition, anomaly detection
Systems
Python, SLURM, distributed multi-GPU training, Docker, Git, FastAPI, AWS, Google Cloud, Microsoft Azure