National University of Singapore
Postdoc under [Kostya Novoselov](https://en.wikipedia.org/wiki/Konstantin_Novoselov)
2022–present
Foresight-Phys: papers into forecastable data
- Built an extraction pipeline that ingests scientific PDFs from public URLs and parses them into structured, queryable JSON — LLM structured-output schemas plus an automatic suitability-filtering pass
- Engineered for throughput, correctness, and observability --- parallel model calls with retry/backoff, a concurrency-safe shared cache, and Langfuse tracing capturing the full experiment lineage (input, predictions, metrics, corrections)
- Scored predictions with proper scoring rules (CRPS, Brier) and calibrated per-field uncertainty; released as the Foresight-Phys benchmark
AI4X conference operations, rebuilt as an engineering system
- Turned the Program Chair appointment — normally hand-run administration — into an engineering system: a pipeline orchestrating deep-research and LLM APIs that reads the open web and returns structured speaker profiles, with routing, fallback, and tracing
- Built embedding-based matching and CP-SAT scheduling to place 100+ speakers into ranked, thematically-coherent sessions
Wyckoff Transformer, a generative model for materials
- Cut inference to 0.05 GPU-ms/structure — 4 orders of magnitude faster than prior models
- Designed a coordinate-free, permutation-invariant autoregressive Transformer encoder that represents crystals by symmetry, improving materials-discovery yield by 1.24×
- Implemented the model in PyTorch; shipped as a production-ready package on PyPI with GitHub Actions CI, YAML-config-driven experiments, and reproducible evaluation for all paper metrics
Leadership
- Program Chair for the AI4X 2025/6 conferences
- Main organiser of the ICLR 2025 Workshop on Machine Learning Multiscale Processes
- Co-PI of a US$3.4m AI Singapore grant on multiscale machine learning
LLM Agent Safety & Alignment
ICML 2026 Workshop, ICML 2026 Workshop, GitHub
- Conceived and developed a novel geometric framework to predict multi-step reasoning failures in LLM agents using semantic abstraction trajectories
- Designed uncertainty quantification (UQ) algorithms for LLM reasoning loops, improving agent error detection and calibration
- Developed and open-sourced reference-checker, a tool for verifying claims and citation references in scientific papers to detect and mitigate LLM hallucinations
- Authored a position paper on Pluralistic Alignment, advocating for aligning AI models with aspirational human values rather than replicating behavioral flaws
Wyckoff Transformer, a generative model for materials
- Developed a coordinate-free permutation-invariant autoregressive Transformer encoder model for material generation and property prediction based on symmetry inductive bias
- Increased materials discovery yield by 1.24×
- Cut inference to 0.05 GPU-ms/structure — 4 orders of magnitude faster than prior models
- Implemented the model in PyTorch; shipped as a production-ready package on PyPI with GitHub Actions CI, YAML-config-driven experiments, and reproducible evaluation code for all paper metrics
- Orchestrated large-scale (10k structures) DFT computations with VASP on an HPC cluster
- Led a team of 6
Wyckoff Transformer, a generative model for materials
- Alignment by design: enforced physical principles directly at the architectural level as hard constraints, mathematically guaranteeing plausible generations
- Solved a systemic failure mode where generative models default to asymmetric outputs despite symmetric training data, mitigating this structural bias via coordinate-free representations
- Increased materials discovery yield (S.U.N.) by 1.24× while cutting inference time by 4 orders of magnitude (to 0.05 GPU-ms per structure)
- Orchestrated large-scale (10k structures) DFT computation loops to verify out-of-distribution stability; led a team of 6
ML for defects in 2D crystals
npj 2D Mater. Appl., npj Comput. Mater., patent, GitHub
- Conceived and implemented a novel sparse representation of crystals with defects
- Cut energy-prediction error 3.7× vs. the best prior model
- Optimised training and inference: 4× less memory, 8× fewer GPU operations
- Built a reproducible data-preparation pipeline: DVC-tracked versioning from raw data through feature extraction to the train/test splits, W&B experiment tracking with per-fold config logging, containerised environments (Docker+Poetry)
- Led a team of 3
Cross-model evaluation benchmark for crystal generative models
NeurIPS 2025 Workshop, dataset
- Created and published an open benchmark dataset spanning 9 generative models for crystal structure prediction --- solicited and curated contributions from the authors of each model (DiffCSP, WyCryst, CrystalFormer, MatterGen, and others) across research groups into a common format, and computed DFT validation for a subset
- Co-authored the LeMat-GenBench benchmark paper (NeurIPS 2025 Workshop), contributing the dataset and analysis of evaluation metrics for comparing generative models of crystalline materials