From research questions to models, benchmarks, and working tools.
01 / Research into practice
01Apodex · May 2026–present
Built pretraining data infrastructure from scratch, evolving on-demand fetching into offline batch processing and integrating domain quality assessment with lightweight URL scoring to guide acquisition.
An acquisition pipeline covering raw HTML, PDFs, and other information-rich documents, with quality scoring to guide collection.
02Apodex · May 2025–May 2026
Built the agent harness from scratch and led its core architecture as development grew to four or five contributors. Designed multi-source verification, parallel tool use, and multi-agent coordination.
A model and workbench for complex research, alongside the open-source FrontierAgent harness.
03I²R, A*STAR · 2023–2025
Led data preparation and evaluation, and co-led model training within a six-person team at A*STAR I²R.
MERaLiON AudioLLM, presented at ACL 2025 System Demonstrations, as part of Singapore’s National Multimodal LLM Programme.
Models and frameworks I contributed to in earlier work.
Verification-focused agents for deep research. Open model: Apodex-1.0-mini.
Earlier work on open-source research agents for long-horizon reasoning and tool use.
Earlier work on a reproducible framework for agentic LLM research.
A universal benchmark for audio large language models. NAACL 2025.
A multilingual benchmark — cross-lingual alignment and cultural reasoning. NAACL 2024.
Resources I maintain and experiments I have built.