CV | Hyeong-seob Kim

Research Interests

AI problems enterprise clients struggle with in production, including real-time speech translation, document parsing for retrieval, LLM-judge evaluation, and GPU-accelerated machine learning.

Publications

Hyeong-seob Kim

MDI: The Minimum Detectable Improvement of an LLM Judge

Accepted to the NeurIPS 2026 Workshop on Trust-AI-Eval (TAE), non-archival.

Sang-Woo Son, Hyeong-seob Kim^*, Hyeonsang Kim, Hyun-woo Cho, Jinmo Kim

Retrieval-Conditional Parsing Score (RCPS): Choosing Document Parsers by Retrieval, Not by Appearance

Accepted to EMNLP 2026 Industry Track.

* Corresponding author

Hyeong-seob Kim, Sang-Woo Son, Hyun-woo Cho, Hyeonsang Kim, Jinmo Kim

WIGVO: Real-Time Bidirectional Speech Translation over Legacy PSTN Calls via Dual-Session Echo Gating

In Proceedings of ACL 2026 System Demonstrations.

Paper Video

Talks

WIGVO: Real-Time Bidirectional Speech Translation over Legacy PSTN Calls (Invited)

Oral Session II, IWSLT 2026 (co-located with ACL 2026), San Diego, CA, USA. July 3, 2026.

Work Experience

Apr 2026 - Present Full-time	Braincrew AI Research Engineer, Research Engineering Team Enterprise AI agent specialist providing AX consulting and AI solution development for Hyundai AutoEver (Hyundai Motor Group), SK hynix, SK AX, and SK Telecom. Design enterprise agent platforms: multi-agent orchestration over ERP APIs behind permission-aware tool gateways, declarative agent runtimes with human-in-the-loop checkpoints, and self-evolving agent loops gated by LLM-as-judge scoring and human approval, for on-premises enterprise deployment. Build production agent components on LangGraph, LangChain 1.x, and DeepAgents: evidence-cited scoring and review agents with citation gates and self-correction middleware, approval interrupts, cross-vendor model fallback, and end-to-end Langfuse tracing and evaluation. Lead technical reviews and evaluation design for client AI systems: versioned tool-interface contracts, independent gold-set evaluation, guardrails against silent approvals, model-swap re-validation, and layered evaluation for LLM agents driving MCP-connected simulators in automotive software verification.
Jan 2026 - Present Research affiliate	RAPIDS LAB (MODULABS) AI Researcher, GPU-Accelerated Laboratory Research on an open-source end-to-end GPU-based AutoML system built on NVIDIA RAPIDS (NVTabular, Ray, GPU XGBoost/LightGBM/CatBoost), benchmarked on OpenML datasets against AutoGluon and LightAutoML.
Jan 2026 - Present Self-employed	WIGTN Founder & CEO, AI Company in Korea Lead WIGTN's research program on real-time speech translation, document parsing for retrieval, and LLM-judge evaluation. All papers are listed under Publications. Staff enterprise AX projects with WIGTN engineers under a delivery partnership with Braincrew, extending the company from research into production delivery. Released WigtnOCR (VLM parsing model for Korean public documents), WIGTN-Coding (an agent harness for Claude Code and Codex), and LLM Loadtester as open source.
Jul 2025 - Mar 2026	Soundmind AI Research Engineer, AX Division Led AI product R&D, building LLM-powered products across RAG, document intelligence, and agentic workflows, from research and prototyping to deployment. Solely designed and built the Soundmind AI Ecosystem, a platform that generates and deploys customer-specific RAG pipelines (AI Console + AI Platform), cutting new engine deployment lead time from 2 weeks to 5 minutes with one-click Docker deployment. Built a RAGAS-based LLM-as-a-Judge evaluation system for public sector RAG (81 Q&A pairs, 3,678 pages) and quantified the performance contribution of each pipeline stage through ablation. Operated an MSA of 9 projects, 161 APIs, and 15+ Docker containers running up to 99 concurrent pipelines, monitored with Grafana and Loki.
Jan 2022 - Sep 2024	Hyundai Engineering & Construction Architectural Site Engineer, Housing Division Led the on-site field verification of H-Silent, a floor noise insulation technology, which earned the first Grade 1 certification from the Korean Ministry of Land, Infrastructure and Transport. Reviewed architectural and structural design changes to remove interference and risk before construction, and managed on-site quality control, workability review, and scheduling.
Jul 2016 - Jan 2022	Dongkuk Construction Construction Project Manager, Head Office Managed contracts, cost, and P&L for RC structure projects, supported new project bidding and estimation, and coordinated administration between head office and sites. Key projects: Doosan Group Bundang headquarters, CESCO Members City, Yonsei University Y-IBS, MBC Sangam parking building.
Aug 2014 - Jul 2016	Doosan Engineering & Construction Architectural Site Engineer, Architecture Division Managed architectural construction schedules and resources, performed quality control against design documents, and contributed to an excellence award at the 2015 LH Construction VE competition.

Selected Projects

Research · Open Source

WigtnOCR: VLM Parsing Model for Korean Public Documents

LoRA fine-tuned Qwen3-VL-2B that surpasses a 15x larger 30B teacher model: 1st in Table TEDS (0.649) on OmniDocBench (CVPR 2025) and best Hit@1 / MRR@10 among 6 parsers on KoGovDoc retrieval.

HuggingFace

Product · B2B · Sole Builder

Soundmind AI Ecosystem: AI Sales Enablement Platform

Sole builder from design to production. Turns customer documents into a deployed, demo-ready RAG PoC in 5 minutes, replacing a 2-week engineering cycle. Dual-LLM document profiling selects the pipeline strategy automatically. An MSA of 9 services and 161 APIs runs up to 99 isolated customer pipelines concurrently.

Details

Hackathons

Snowflake AI & Data Hackathon 2026 Korea · 2nd Place, Tech Track

WIGTN FLAKE: Purpose-Driven Neighborhood Intelligence Platform

An AI expert panel cross-analyzes real estate, foot traffic, card spending, and telecom data to recommend locations by purpose. Snowflake Cortex and GPT-4o hybrid orchestration.

Video News

Google Cloud Rapid Agent Hackathon

Custos: Always-On AI Code Review Bot for GitLab

Four reviewer personas analyze merge requests in under 60 seconds, while an hourly proactive lane files prioritized issues and recalibrates from maintainer reactions. Google ADK, Vertex AI Gemini, Cloud Run.

Devpost Video

H0: Hack the Zero Stack · Vercel × AWS

OpenSlot: Zero-Oversell Ticketing Infrastructure

Each seat is a row in a multi-region Amazon Aurora DSQL ledger with versioned conditional updates, verified against a simulated 2,000-buyer stampede with zero oversells. Next.js 15 on Vercel.

Devpost Video

Gemini Live Agent Challenge · Google

TimeLens: Multimodal AI Museum Curator

Voice and camera AI curator processing audio and vision in real time on a Gemini Live API and ADK dual pipeline.

Devpost Video

Education

Jun 2023 - Feb 2024	National Institute for Lifelong Education Bachelor's degree in Architectural Engineering
Mar 2009 - Aug 2015	Kyungbok University Associate degree in Architectural Design

Honors & Awards

Apr 2026	2nd Place, Tech Track, Snowflake AI & Data Hackathon 2026 Korea Team WIGTN, WIGTN FLAKE
Dec 2022	Outstanding Employee Award, Technological Competitiveness Hyundai Engineering & Construction
Apr 2015	Honorable Mention, 11th Excellent Graduation Thesis Exhibition Architectural Institute of Korea

Certifications

Mar 2026	Fundamentals of Accelerated Data Science NVIDIA Deep Learning Institute
Dec 2025	Engineer Big Data Analysis (빅데이터분석기사) Korea Data Agency