Naver Search US / Generative AI Multimodal team- Designed and implemented the advanced AI search engine--interactive UI delivering the answer from the search result--using various customized cutting-edge LLMs fine-tuned on the distributed environments (i.e., Kubernetes)- Implemented scalable APIs and microservices to seamlessly integrate LLM technologies with diverse web platforms, enhancing interoperability and enabling real-time data processing and analytics.- Led the cross-functional teams of ML scientists, ML engineers, and software developers to build and optimize AI-driven solutions, ensuring alignment with strategic goals and delivering projects on schedule.- Utilized the techniques such as transfer learning, hyperparameter tuning, and model pruning to optimize LLM performance and systematic approaches such as dynamic batching, and paged attention. - Developed customer-centric, bespoke AI solutions for customers by leveraging LLM capabilities to enhance user experience, automate complex tasks, and provide actionable insights.- Drove innovation by exploring emerging LLM technologies and integrating them into business strategies, fostering a culture of continuous improvement and technological advancement.

Alexa Ranking & Arbitration Science teamDesigned/implemented the deep learning models and systems to deliver tailored customer experiences for the enterprises partners (EP) to empower the enterprise-designed user scenarios and their own virtual agents (like Alexa) in addition to all Alexa capabilities with the consideration of contexts such as device location, time, user subscription plan, etc.Achievement:[Enterprise Program]- Redefined the problem of predicting relevant Alexa domains/3P skills for given user utterance as that of predicting relevant Alexa domain/3P skill groups additionally for given relationship with enterprise partners.- Re-designed the two-phase DL models to take ensemble methods with considering minimized calibration, agility, robustness, and guardrails, along with the context modeling for EP-defined scenarios.- Led the agreement from the stakeholders including two scientist teams and four engineering teams.- Led scientist/engineering teams to design/implement/migrate to the new runtime system for serving the new model architectures.- Designed/implemented the automatic onboarding system and the necessary toolings for the business team and onboarded 5+ enterprise partners.- Implemented the full modeling lifecycle from continuous influx of historical data to release the best model to production.[Recall Improvement Project]- Initiated to re-architecture to address the recall issue on developer-provided data caused by the data distribution gap.- Designed/implemented the graph-based model to quickly capture updates from developer and to scale to tens of thousands of Alexa skills regardless of their types.- Designed/implemented the inference code on the Alexa platform.- Led the agreement from the stakeholders--scientist/engineering team--with the experiment results from prototype.- Led the scientist/engineering teams to implement and deliver the product

Alexa Ranking and Arbitration Science teamDesigned/Implemented the heterogeneous routing ML system composed of two stages using deep learning models for customer's utterances to route toward the appropriate Alexa domains and skills with the consideration of rich contextual information.Achievement:- Designed/built the deep learning model using DyNet library for routing and arbitrating thousands of third party developed Alexa skills, which is currently serving the production traffic in seven different locales including non-English language such as Japanese, German as well as English using in different regions (e.g., Canada, UK, IN).- Designed/Implemented the entire offline pipeline from scratch for data collection, model build, evaluation, and release to the production system.- Designed the model artifacts (structure/interface) used in runtime system.- Implemented the core inference engine in the production system using JNI interface integrated with DL runtime library written in C++ (DNNRT, internal library).- Implemented a variant of LSTM (LSTM w/ peephole connection) missed in the runtime library written in C++.- Implemented the model converter from Dynet to DNNRT library in Python from scratch.- Implemented the model converter from PyTorch to DNNRT library in Python.- Implemented Pyspark jobs for collecting and processing diverse source of data.- Designed the next generation of inference engine in production to address the pain-points coming from complex components and external dependencies on different services.- Designed the ML model architecture to accommodate private Alexa skills and released for experimentation in shadow mode.- Designed the ML model architecture and experimentation plans for arbitration between Alexa-own domains and third-party developed skills.

AI & Research Group / Knowledge & Conversation NL- Designs the architecture of the runtime workflow for language understanding using deep learning.- Implements the runtime component and API of deep learning models for language understanding.- Builds deep learning models for Cortana in speaker and new skills.

Community Shopping Department / Zebra team- Designed an entire system for customer reviews from backend storage to user interface.- Analyzed resources/services in terms of capability, limitation, and cost.- Took into consideration various vulnerabilities—the abuse prevention, failure tolerance.

Item and Offer Pipeline Department / Search Data Aggregation (SDA) team- Designed/implemented search index data used in ranking and visibility of items on the search results.- Designed/implemented the business logics under the Complex Event Processing (CEP) environment.- Fixed data consistency issues under high throughput (in billions/day), multi-threaded settings.- Optimized the business logics to reduce latency and to suppress unnecessary traffic.- Improved manual/periodic operations to reduce operational loads.

- Designed the home cloud system with consideration for easy setup, security, expansibility, and availability.- Implemented the hub of home cloud system as Android application.- Incorporated the connectivity of various home devices--computer, smartphone, tablet, storage (NAS), and surveillance camera.- Incorporated security concerns—access-control for user, secure data channel (SSH tunnel).

- Documented the comparison of UI’s of two smart phones running on Windows Mobile 6.5.- Analyzed the anticipated threats on smartphones based on the case study of security incidents on Symbian and WIPI in the past.- Analyzed the structure of Fixed Mobile Convergence (FMC) network with its potential vulnerabilities.

- Compared user interfaces of two smart phones using Windows Mobile 6.5- Wrote a report on the result of the comparison and the further improvement- Collected past incidents about malicious software on Korean mobile platforms--Symbian, WIPI- Analyzed the potential attacks by malicious code on Iphone OS, Android, and Windows Mobile- Wrote a report on the future threats and their solutions- Studied the structure of Fixed Mobile Convergence(FMC) network- Analyzed the potential threats on the FMC network- Wrote a report on the vulnerabilities and their solutions for each threat

- Debugged open source video conference library supporting H.323 and SIP called OPAL and OpenPhone- Modified X.264 codec to carry an extra text information to the header of Network Abstraction Layer(NAL) unit- Made H.264 decoder extract the extra data from modified NAL unit

- Maintained computer network and update website- Ran Samba server on Unix to share network disks and printers