Đinh Thiên Ân

Zero-shot, two-stage pipeline for document-level entity and relation extraction. First, entities are extracted by three heterogeneous LLMs (DeepSeek-R1 distill, Llama-3.3-70B, Qwen-2.5-32B) and consolidated through deduplication and majority-vote typing. Then relations are extracted with prompts that constrain head and tail to the consolidated entity set. No domain-specific training is required.

Interactive search system for finding events in a large video archive. The index covers 1,487 videos in 125 groups and 524,891 keyframes (news, CCTV traffic cameras, cycling). Features: Vietnamese text-to-image visual search on SigLIP2 with LLM query translation, OCR text search, ASR transcript search, object-detection filtering, temporal multi-event search, voice queries, and a built-in DRES submission bar.

Searches, gathers and analyzes online information (news, social media, public data) about customers from their personal or business identification details. It builds a risk blacklist from crawled news with LLM-based extraction and risk labeling, and an agent backend matches queried names and returns risk information and scores.

System for judging which of two stories is narratively closer to an anchor (abstract theme, course of action, outcome), using contrastive fine-tuning of sentence transformers trained on synthetic data. Character names are pseudonymized with consistent placeholders so the model learns narrative structure rather than names.

Track 2 (Transportation Safety Understanding and Captioning, Sim2Real) asks for detailed pedestrian and vehicle captions across five phases of each traffic-safety scenario, plus multiple-choice VQA. Models may be trained only on the synthetic SynWTS digital twin and are evaluated on real WTS videos. LoRA-fine-tuned InternVL2.5-8B with global scene frames plus local crops around pedestrian and vehicle boxes.

Classifies Vietnamese social-media posts into sarcasm in text only, in image only, in both, or none. Two approaches: ViSoBERT (text) + BEiT (image) fusion, and an end-to-end fine-tuned CLIP with a custom multimodal classifier (focal loss with class weights, mixed precision). Predictions combined by voting.

A manager agent decomposes a user request, discovers worker agents (Researcher, Designer) through an on-chain registry, negotiates prices with LLM reasoning, locks funds in escrow, validates the deliverables and releases payment gaslessly. Agents mint NFT 'passports' (DIDs). A React observer dashboard streams the whole process live.