Clearly Local Partners with KuCoin Boost AI Translation Quality by 20% Across Key Global Markets
Leading global cryptocurrency exchange platform KuCoin recently partnered with Clearly Local to refine its AI translation prompts for Turkish, Vietnamese, and Brazilian Portuguese—languages chosen for their relevance to key markets and distinct linguistic characteristics. The collaboration enhanced translation quality across all three languages by 20%, reduced prompt length to one-third, and provided actionable insights for KuCoin to improve its AI-generated translations for other languages.
A Need to Improve Multilingual Output Accuracy
In just two years, many organizations have been experimenting with integrating LLMs (large language models) into their translation workflows. A key challenge in this shift has been standardizing output quality to match the relative reliability of traditional NMT (neural machine translation) solutions.
KuCoin encountered this issue firsthand when using ChatGPT. Initially, the company tried improving translations by developing custom prompts in-house. However, the results fell short of expectations and failed to meet users’ quality standards. Lacking the linguistic expertise and technical knowledge to fix the issue, KuCoin turned to external support.
It sought a localization partner with expertise in AI-driven solutions and a network of professional linguists to deliver strategic insights. “Clearly Local was the obvious choice,” said Adam Yuan, Head of Localization at KuCoin. “We’ve collaborated with them before and value how they seamlessly combine cutting-edge technology with human-driven insights.”
Setting the Stage for Success
Clearly Local’s team quickly met with KuCoin to align on business goals and success metrics. After evaluating the challenges, their project managers, solutions architects, and AI experts identified human evaluation metrics as the most effective way to assess and compare the performance of LLM-generated translations.
Specifically, they implemented an approach that combined qualitative linguist feedback, including an overall quality score, and binary assessment (Yes/No), which assesses whether issues in the previous round have been fixed. These two evaluation methods helped to clearly define if quality standards were met.
An example of both evaluation methods used in the project.
To streamline the process, the team built the Observable HQ Notebook—a centralized hub for up-to-date locale-specific data and quality benchmarks—ensuring prompt consistency and accuracy.
They then began recruiting linguists specializing in the three target languages. Interestingly, during this process, several candidates declined to participate, voicing ethical concerns about supporting AI tools they believed could undermine human translators. Evan Zheng, the project lead, encapsulated this resistance while recounting one linguist’s blunt refusal: “I won’t contribute to tools that threaten our profession.”
Next, the team did a preliminary review of KuCoin’s existing prompts and translations, which were based on an English article from KuCoin’s website. During the review, they immediately noticed the lengthy glossary explanations from KuCoin’s term base, and hypothesized whether removing them could improve the current output quality. The logic was simple to their solutions architects: verbose explanations dilute prompt focus, effectiveness, and translation quality. By refining the prompts into cleaner, more streamlined versions, could they address these issues?
After reviewing the original content, the team developed a three-step prompt optimization strategy:
- Establish a baseline reference: Linguists provide feedback and assign an overall quality score to the output of both prompts (one version with glossary explanations and one without).
- Iterate and refine: Starting with the highest-scoring prompt version, conduct another three rounds of adjustments based on the same scoring method, but introduce binary assessment to confirm the resolution of prior issues.
- Finalize optimization: Lock in the highest-performing prompts.
Beyond the Prompt: A Structured, Human-Centric Approach
Round 1: Simplicity Prevails
In the first round of optimization, Clearly Local’s linguists assessed the output of both versions of KuCoin’s original prompts: one with glossary explanations and one without, and provided feedback about common issues in the output as well as an overall quality score for each.
Their assessment was unanimous: the streamlined prompts (without glossary explanations) produced significantly higher-quality translations across all three languages. Furthermore, linguists reported that removing the glossary term descriptions did not have a noticeable effect on terminology usage between the two translations.
Left: V1 (prompt with glossary descriptions) Right: V1.5 (prompt without glossary descriptions)
Rounds 2 and 3: A Sisyphean Task
The second round showed an overall quality improvement of 10% for all three languages, which was encouraging. But by the third round, a critical pattern started to emerge: even with better prompts, new and recurring errors kept emerging. This highlighted a core limitation of ChatGPT and similar AI tools—their unpredictability. Unlike rule-based systems, these models can’t consistently deliver standardized outputs.
An example from the Turkish language score card showing numerous recurring issues from round 3.
For instance, Anna Jin, Clearly Local’s Solutions Architect Director, found that ChatGPT occasionally skipped entire text sections, particularly in lengthy content. Similarly, the Brazilian Portuguese linguist reported untranslated segments in some files, which the team had to fix manually to ensure accuracy.
Clearly Local’s solutions architects identified limited training data as part of the problem. Due to budget and time constraints, the team relied on a single English article from KuCoin as their only training input. They repeatedly refined the AI’s prompts to improve translations of that specific article—fixing typos and awkward phrasing. However, over time, the AI grew overly specialized: it excelled at translating that one article but struggled with new, unfamiliar content.
This is a classic case of overfitting, a common AI training pitfall. Overfitting happens when a model becomes too tailored to a narrow task (e.g., translating one article perfectly) and loses its ability to adapt to broader tasks (e.g., translating any article perfectly).
In the high-risk crypto industry, where inaccurate content can directly harm customers’ finances and erode trust, Clearly Local’s discovery underscored the urgent need for KuCoin to integrate diverse training data and maintain human oversight in future optimizations.
Round 4: From Flaws to Progress
In the final optimization round, Clearly Local compiled all scores and feedback, tasked linguists with final prompt tweaks, and delivered a unified, optimized prompt.
Elevating AI Localization for KuCoin
Following the implementation of Clearly Local’s optimized AI prompts—which were reduced to one-third of their original length—KuCoin achieved a 20% improvement in overall translation quality. Adam reported a similar increase in quality for other languages, including Traditional Chinese.
Clearly Local’s human-in-the-loop approach proved essential for refining outputs. Through iterative human feedback, the gains achieved were clearly noticeable, as one linguist observed:
“The general quality right now is close to a professional translator, and I believe this is the best achievable result for machine learning. It still has some issues for punctuation, but most of them are gone.”
However, the inevitability of issues like punctuation inconsistencies in AI translation outputs highlight the enduring importance of Clearly Local’s hybrid framework: AI scalability paired with human oversight ensures quality for high-risk industries like crypto, where even minor errors can impact user trust.
Key Recommendations for KuCoin’s Future AI Strategy
The project revealed critical lessons about AI’s limitations and opportunities. To sustain progress, Clearly Local proposed these actionable strategies for KuCoin’s future AI initiatives:
- Incorporate diverse data in future iterations: Integrate varied linguistic inputs to help the AI generalize effectively, avoiding over-reliance on narrow datasets.
- Streamline refinement cycles: Limit optimization efforts to 2–3 rounds per project, balancing precision with adaptability for scalable results.
- Restructure team roles: Assign linguists distinct responsibilities (prompt refinement vs. output evaluation) to ensure objectivity and reduce bias.
- Sustain human-AI collaboration: Maintain human oversight to address nuances (e.g., punctuation) and contextual challenges that technology alone cannot resolve.
These insights revealed a hard truth about AI-driven translation: without human expertise, even advanced models risk repeating costly errors.
Partnering with Clearly Local enables KuCoin to deliver seamless, localized crypto trading experiences—just as KuCoin empowers users worldwide to securely access global markets and navigate the complexities of cross-border transactions. By integrating Clearly Local’s human-guided AI strategy, KuCoin has solidified its position as a leading cryptocurrency exchange.
