I enjoy designing and refining AI pipelines, whether it's building RAG systems, orchestrating AI agents, or optimizing end-to-end workflows for production.

• Building B2B production AI systems end to end at Braincrew (Research Engineering Team)
• Running WIGTN, an AI research group: speech translation, document AI, developer tools
• Researching GPU-accelerated ML at RAPIDS LAB (MODULABS): study notes · WikiDocs

Full career details: hyeongseob91.github.io

• WIGVO: Real-time bidirectional speech translation over legacy PSTN calls (Published in ACL 2026 System Demos, 1st Author)
• WigtnOCR-RADP: Retrieval-Aware Document Parsing. Proposes RCPS, a retrieval-grounded metric for choosing document parsers by retrieval rather than appearance, developed from a failure mode found while building WigtnOCR-v1 (EMNLP 2026 Industry Track, Under Review)
• WigtnOCR-v1: VLM-based SLM for Korean public document parsing. A 2B model trained on data distilled from a 30B teacher and quality-filtered by a 122B judge (top 75% retained), surpassing the teacher. Released with the KoGovDoc-Bench dataset

• WIGTN-Coding: Claude Code plugins for an AI-Native development workflow
• LLM Loadtester: LLM serving performance benchmarking tool (TTFT/TPOT/Goodput visualization)

An AI research group founded in 2026. WIGTN takes research from idea to working system: real-time speech translation published at ACL 2026, VLM document parsing models released on HuggingFace, and open-source developer tooling. The group also serves as a test partner for companies validating new AI products, and the same team builds together at hackathons.

• 🥈 Snowflake AI & Data Hackathon 2026 Korea · Tech Track 2nd Place
  WIGTN FLAKE: A purpose-driven local intelligence platform. Users pick what they want to do (open a cafe, run targeted rental marketing, place outdoor ads, invest in real estate) and an AI expert panel cross-reads four datasets (real estate, foot traffic, card sales, telecom contracts) to recommend the best neighborhoods, surface anomaly signals before they show up in headlines, and project six-month trends. Built on a hybrid Snowflake Cortex × GPT-4o orchestration that lets the analysis read like a conversation between domain experts rather than a dashboard.
• Custos: Always-on AI code review bot for GitLab. Four reviewer personas review merge requests in under 60 seconds, while an hourly proactive lane files prioritized issues. Built on Google ADK, Vertex AI Gemini, and Cloud Run (Google Cloud Rapid Agent Hackathon)
• 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 (Vercel × AWS · H0: Hack the Zero Stack)
• TimeLens: Multimodal AI museum curator processing voice + camera in real time on a Gemini Live API + ADK dual pipeline (Gemini Live Agent Challenge)