Algorithm Engineer at Reolink Singapore, where I build production RAG pipelines and agentic systems, from fine-tuned retrieval and on-device inference to the evaluation harnesses that tell you when things quietly break. I hold a Ph.D. in Computer Science from NTU Singapore, with research roots in audio-language models and cross-modal learning. I write here about the interesting problems and challenges I run into along the way.
Recent writing
- In praise of a lazy coding agent: Codex vs Claude Code
- Don't trust the LLM to do math: a placeholder pattern for production RAG
- Semantic Deduplication for RAG: Reducing Redundancy in Data
Selected work
- AI Invoice & Expense Claims Automation Telegram bot that reads supplier invoices and staff receipts with Claude, then files them to Drive and Sheets for an F&B operator.
- SakuraSensei: Japanese Conversational AI Tutor Japanese tutor bot with persistent memory, RAG over JLPT and dictionary data, and cloze drills generated from YouTube audio.
- Audio Captioning with Transfer Learning and Latent Space Regularization Transfer learning plus a reconstruction latent space similarity regularizer for automated audio captioning.
- Acoustic Event Detection for Emergency Telephony Streaming PyTorch model that detects nine acoustic event classes in emergency call audio, with caption output and Docker deployment.