All posts
How Surf built Korea's first AI-centric trading experience with Bithumb
Surf and Bithumb Launch Korea's First AI-Centric Trading Experience Bithumb users can now research assets with real-time digital asset AI analytics and execute trades in a single app. Surf and Bithumb, one of Korea's largest digital asset exchanges, today announced a strategic partnership through which Surf's AI intelligence layer now runs inside the Bithumb app, giving Bithumb's more than 8 million users Korea's first AI-centric trading experience, with live analytics and trade execution in one place.
AnnouncementSep 1, 20265 min read
Bithumb, one of Korea's largest cryptocurrency exchanges with more than 8 million users, set out to become the country's first AI-centric exchange. Surf was selected as the core development partner over frontier AI labs to build an intelligence pipeline that covers all 477 Bithumb KRW markets, delivering the conviction of three hours of manual research in seconds to all users on Bithumb.
The Challenge
Trading has always rewarded whoever adopted the most advanced technology first, from the telegraph to the Bloomberg terminal. AI is the next technology traders are rushing to adopt: leading exchanges such as Binance, Coinbase, OKX, and Bybit have shipped AI-powered features, but no major Korean exchange had made AI the core investment experience. Bithumb saw the opening—and its ambition was not another chatbot bolted onto an exchange.
The requirement behind every AI-driven insight on a user's screen was the same: information that refreshes with the market, in native Korean, with zero tolerance for error, while staying perfectly compliant with local and business regulations. Off-the-shelf AI can get to an answer that sounds correct, given enough time and attempts. But for a regulated exchange responsible for millions of traders' money, "usually sounds correct" is not a deployable standard.
Neither better prompting, better models, nor more data alone would meet that standard. The answer was to own the feedback loop around the model: the harness.
The Solution
To address constraints that no amount of prompt engineering on a general-purpose model can account for, Surf AI focused on the custom harness instead of the model itself: the guardrails and quality-control system wrapped around the AI, governing what goes in and verifying everything that comes out.
Instead of letting the model search for its own context (which often leads to slower throughput and hallucination), Surf's harness prepares it for the model every 5 minutes: market, on-chain, news, and social data are frozen into a single snapshot, screened through more than 100 quality checks, and distilled into a handful of clear, explainable signals.
However, no matter how well prepared the input (data & prompt) is, raw LLM output still carries quality concerns in how generated insights are worded and presented. To ensure content consistency, our post-generation evaluation layer deterministically hard-blocks any output containing financial advice or competitor mentions, and checks every figure against source data.
Language fluency is also a bottleneck for frontier LLMs: the majority of their training is done in English, causing unnatural Korean output. We designed the system so content is generated in English first, then translated and quality-checked by an agentic evaluator trained on human labels from Bithumb and Surf's Korean reviewers. Every flagged failure is logged and shared with Bithumb's engineering team, so the system improves with each production cycle.
To ensure uptime and keep insights chronologically consistent, we designed 3-way fallback routing: First, when the system flags a quality concern in the LLM-generated output, the content is automatically regenerated with the exact mistake spelled out in the next prompt. Second, if a model provider goes down at any point, the pipeline seamlessly switches to a backup model from a different vendor, resuming from where it stopped. Lastly, anything that cannot be fixed is blocked entirely: users only ever see complete and chronologically aligned information.
Results
The result is simple: the conviction that took hours of manual research is delivered in seconds to Bithumb's 2 million daily active users. 200 tokens are scored, analyzed, and written up into reports using the most recent data, along with insights on how traders can act. This flow funnels user trading activity back to Bithumb.
Factually accurate and compliant output that sounds like a local
Across 2.5 million generation and validation attempts, the system intercepted and rerouted 300K unsuccessful executions instead of letting them pass silently. Backed by a natively trained LLM translation judge with a 300+ term institutional glossary, Korean quality reached 95% first-pass and 99% second-pass approval. (add how report + daily brief is benefiting here)
Efficient throughput
The harness design controls the tasks given to the model and the procedures it executes. This prevents the model from looping inefficiently, allowing rapid throughput and a lower token burn rate: 450+ markets refreshable at roughly 0.05-second p95 latency. (AI reccomended assets being refreshed)
99.2% production API availability
Bithumb users never experience server-side errors. No misaligned insights, no partial data, no blank screens during provider outages. Every failure is absorbed before it reaches the interface, so the AI experience stays trustworthy under an exchange-grade SLA. 99.9% probe success across 50K+ live health checks, without Bithumb's team ever firefighting the pipeline.
Metric | Detail |
|---|---|
500M+ | Market and technical data points processed |
2.5M | AI generation and validation runs executed |
750K+ | Structured AI intelligence outputs generated |
3.M+ | Korean quality evaluations performed |
95%+ | Automated Korean quality approval rate |
98%+ | Multi-source snapshot success rate |
450+ Markets | Live coverage across every Bithumb KRW market |
99%+ | Production API availability |
Why Surf?
We saw Surf as the partner best aligned with the direction Bithumb is pursuing. Turning the vast data of the digital asset market into intuitive insights that retail investors can actually digest requires both AI capability and the ability to refine social data, and Surf brings both.Goals & Expectations
Bithumb's goal is to create an environment where users can clearly understand complex market movements through data. We expect the AI signals and analysis delivered through this collaboration to help close the information gap that retail investors face— Bithumb team
To mark the launch, From Aug. 31 through Sept. 30, 2026, All user that joins Bithumb through Surf A's event page will receive a 50,000 KRW bonus completion of their first trade. Users must log in to or create a Surf account to redeem theirs. Learn more
