Portfolio
Things I've built and the research I've worked on — from code-switching speech recognition and deepfake detection to multi-agent LLM systems. Most have code on GitHub.
AutoNxt AI · Kaizenics
A domain-agnostic agentic AI platform — multi-agent orchestration that adapts across domains. Built at AutoNxt AI for Kaizenics, with a software-defined-vehicle application running virtual cars on RemotiveLabs ECUs with Android Automotive and CARLA.
AutoNxt AI · digital.auto
A child-presence safety reminder that helps prevent children from being left unattended in a vehicle — built at AutoNxt AI with digital.auto.
Research @ A*STAR I²R, Singapore
A hybrid Whisper-encoder + LLaMA-decoder ASR system for Malay–English code-switching, trained with Noisy Student Training and teacher–student pseudo-labeling for low-resource robustness.
Reduced WER from 34% → 28% on the CS-ASR benchmark
A modular multi-agent system automating customer support, product recommendation, and dynamic pricing — RAG-based reasoning and memory over a streaming backend with a TypeScript front-end.
End-to-end prototype: ingestion → embeddings → agents → API → UI
Research @ CIS Lab, National Chung Cheng University (Taiwan)
A hybrid CNN + InceptionNeXt + Vision Transformer pipeline for video deepfake detection, trained across Celeb-DF (V2), DFD, DeepfakeTIMIT and WildDeepfake with strong augmentation.
91% accuracy on CelebDF-V2, 94% on DFD
A Transformer encoder–decoder that generates fashion product descriptions, using BEiT for visual features and GPT-2 for generation, evaluated on the FACAD dataset.
ROUGE-L 67.5 / BLEU-1 65.4 — beats ViT & DeiT baselines
A retrieval-augmented generation pipeline for question-answering over financial documents.
Vietnamese sentiment classification experiments comparing different word-embedding approaches.
Large-scale graph analytics with GraphFrames on Apache Spark.