Clober is a DeFi protocol company — “Building DeFi primitives that matter” — building an order-book DEX on-chain. The core problem we worked on was bringing a decentralized order-book exchange fully on-chain and shipping it as an optimized, usable product. Over four years of founding-stage work I helped take Clober from zero to one across the full product surface — protocol internals, backend, and frontend.

Beyond the engineering itself, what four years of founding-stage work at Clober gave me: hands-on zero-to-one product experience (planning, UI/UX, execution, launch); comprehensive domain knowledge in finance; and practical experience in DeFi architecture design and financial modeling. These are the parts I carry forward the most.

The engineering work I personally did during that period breaks down as follows:

Task 1 — Clober v1 (2022.04 – 2023.03) — clober.io

• Built Clober v1 across the full stack (Smart contracts + Backend + Frontend).
• On the contract side, implemented gas-optimized on-chain order matching using segmented segment trees and octopus heaps.
• Write-up: ethresear.ch post · recap thread · gas benchmarks
  
  

Task 2 — DEX aggregator on Polygon zkEVM (2023.03 – 2023.06)

• Built a DEX aggregator for Polygon zkEVM in Rust, solo development, to support Clober’s growth on the chain.
  
  

Task 3 — Coupon Finance (2023.06 – 2024.06) — coupon.finance

• Built the frontend and backend of Coupon Finance, a sister protocol built to drive Clober’s growth.
• Coupon Finance is a fixed-rate DeFi lending protocol that tokenizes interest rates as tradable “coupons” on an on-chain order book — collapsing the deposit–borrow APY spread down to the order book’s bid–ask spread, so both lenders and borrowers get tighter rates than on traditional money markets.
  
  

Task 4 — ETH/USDC vault strategy (2024.06 – Present) — base.clober.io

• Responsible for the development and management of Clober’s ETH/USDC vault strategy, end-to-end.
• Improved market-making strategy using Markov modeling.
• Built the data pipeline powering strategy backtesting and simulation.
• Designed a real-time on-chain event tracking pipeline (Ponder-style indexing → Grafana dashboards) for live strategy monitoring and data analysis.
• Optimized DEX ↔ Clober arbitrage to reduce market-maker losses.
• Launches: Monad mainnet (2025.11), Base chain vault (2026.01).

Upstage AI is a Korean AI company focused on enterprise AI. I joined as a research engineer on the Document AI team during its early phase, working on the foundational models that later became part of Upstage’s document-understanding product line.

Task 1: LayoutLM training and fine-tuning

• Trained and fine-tuned LayoutLM — a multimodal transformer that jointly models page layout and text — for Korean document understanding.

Task 2: OCR labeling dataset

• Built the labeling dataset used to train the document understanding models.

Task 3: Document parser

• Built a parser for document processing.

Brunel AI is a startup that builds AI-powered search products for patents. This was a formative experience in thinking about AI from a product perspective — the models I trained shipped into the actual product and came back with real user feedback.

Task: AI engineering across the stack

• Helped scope research problems into AI tasks that could be realized as real product features (e.g., deciding which data to collect for long-term value, and framing user problems as concrete ML tasks).
• Improved search quality by developing a model that recommends relevant patents from a user’s input query.
• Built the supporting infrastructure for the data pipeline and model serving.

NAVER Clova AI is one of Korea’s leading AI research organizations. I worked in the LaRva team on Korean language modeling, which is where I built deep practical experience with large-scale training on distributed GPU clusters. In six months I was able to contribute to several tasks at once, thanks to the team’s resources and mentorship.

Task 1: Pretraining large-scale language models (BERT, RoBERTa) in a distributed GPU environment

• Trained on up to 64 V100 GPUs with large-scale corpora in a distributed setup.
• Ran hundreds of pre-training experiments and learned how to avoid CPU/GPU bottlenecks, use mixed precision effectively, and manage large training datasets.
  
  

Task 2: Efficient and lightweight pretrained language models

• Studied how to apply Lample’s Product Key Memory (PKM) to pretrained LMs while avoiding catastrophic drift during Masked-LM pretraining and downstream fine-tuning. The work was accepted to Findings of EMNLP 2020.
• Applied knowledge distillation to build a lightweight Korean RoBERTa.
  
  

Task 3: KorQuAD 2.0 leaderboard — 1st place (F1/EM: 83.54 / 66.95, 2020-04-22)

• KorQuAD 2.0 is a reading-comprehension dataset in the spirit of Google’s Natural Questions: the model has to find the correct answer span inside a full Wikipedia page, where the answer can be a table, a list, or a paragraph — much harder than single-paragraph QA.

Platfarm is a startup that builds products that recommend emoji based on chat text. This was my first industry role and where I first learned how an engineering team collaborates day to day.

Task: Emoji recommendation from chat text

• Annotated and cleaned a raw chat corpus for training.
• Built a sentiment-analysis-based emoji recommendation model.