Personal website, research publications, and machine learning portfolio of Dongchan Kim — Computer Scientist & Applied Scientist.https://dongchankim.io/https://dongchankim.io/posts/pia-personal-intelligence-agent/https://dongchankim.io/posts/pia-personal-intelligence-agent/Introduces PIA, a personal intelligence agent that solves critical data loss in general LLM conversational memory by transforming medical dialogues into structured clinical records and multi-dimensional longitudinal understanding.Tue, 15 Sep 2026 00:00:00 GMThttps://dongchankim.io/posts/retrieval-collapses-when-ai-pollutes-the-web/https://dongchankim.io/posts/retrieval-collapses-when-ai-pollutes-the-web/Pioneering empirical study published at TheWebConf (WWW 2026) identifying 'Retrieval Collapse'—an ecosystem-level failure mode where AI-generated web pollution degrades external evidence discovery in RAG pipelines and search engines.Wed, 18 Feb 2026 00:00:00 GMThttps://dongchankim.io/posts/curated-deep-learning-resources/https://dongchankim.io/posts/curated-deep-learning-resources/A curated collection of essential textbooks, interactive courses, and foundational papers for mastering modern deep learning and neural architectures.Wed, 10 Jan 2024 09:00:00 GMThttps://dongchankim.io/posts/skill-shortlister-patent/https://dongchankim.io/posts/skill-shortlister-patent/US Patent 11,568,863 B1 granted to Amazon Technologies. Scalable neural shortlisting architecture that prunes massive skill spaces into candidate shortlists for low-latency voice assistants.Tue, 31 Jan 2023 09:00:00 GMThttps://dongchankim.io/posts/coupled-representation-learning/https://dongchankim.io/posts/coupled-representation-learning/Published at IEEE SLT 2018. A coupled representation learning framework that models hierarchical inter-dependencies between domains, intents, and slots in spoken dialog systems.Tue, 18 Dec 2018 09:00:00 GMThttps://dongchankim.io/posts/efficient-large-scale-domain-classification/https://dongchankim.io/posts/efficient-large-scale-domain-classification/Published at ACL 2018. A scalable neural architecture incorporating personalized attention for routing utterances across thousands of overlapping domains with 100x training speedup.Sun, 15 Jul 2018 09:00:00 GMThttps://dongchankim.io/posts/scalable-neural-shortlisting-reranking/https://dongchankim.io/posts/scalable-neural-shortlisting-reranking/Published at NAACL 2018. Introduces a two-stage neural shortlisting and reranking approach for ultra large-scale conversational NLU across 1,500+ domains.Fri, 01 Jun 2018 09:00:00 GMThttps://dongchankim.io/posts/adversarial-adaptation-synthetic-stale-data/https://dongchankim.io/posts/adversarial-adaptation-synthetic-stale-data/Published at ACL 2017. A novel adversarial neural domain adaptation framework that bridges distribution mismatch caused by synthetic and stale training data in NLU systems.Sun, 30 Jul 2017 09:00:00 GMThttps://dongchankim.io/posts/domain-attention-ensemble-of-experts/https://dongchankim.io/posts/domain-attention-ensemble-of-experts/Published at ACL 2017. An ensemble-of-experts model with domain attention for adapting conversational systems to new domains without full retraining.Fri, 28 Jul 2017 09:00:00 GMThttps://dongchankim.io/posts/adaptive-data-transmission-cloud/https://dongchankim.io/posts/adaptive-data-transmission-cloud/Published at IEEE/ACM IWQoS 2013. Introduces Adaptive TCP (ATCP), a lightweight end-host protocol prioritizing latency-sensitive short flows over background bulk traffic in cloud data centers.Mon, 24 Jun 2013 09:00:00 GMT