How DeepL Built a Translation Powerhouse with AI with CEO Jarek Kutylowski
Jarek Kutylowski, CEO of DeepL, discusses the company's technical strategy for competing with large language models in the translation space. He covers their focus on specialized model architectures, the critical role of curated data, the engineering challenges of building custom GPU data centers and large-scale inference systems, and the future of AI-driven translation in enterprise workflows.
Model Architecture and Strategy
DeepL’s success stems from focusing on specialized models for translation. When the industry shifted to neural machine translation around 2017, it created an opportunity for startups to innovate. DeepL developed proprietary architectures that were better suited for translation than general-purpose models of the time. The key is balancing two opposing goals:
- Accuracy: The translation must remain faithful to the source text.
- Fluency: The output should sound natural and native in the target language.
This requires a mix of models that are good at both “copying” and “writing.” While DeepL evaluates and uses large pre-trained models as a base, they add significant compute and training on top, using their own highly curated datasets. This specialized approach leads to models that, even if comparable in size to large general-purpose models, are more focused and performant on the specific task of translation.
Data as a Differentiator
A core advantage for DeepL is its long-term investment in building a high-quality, curated dataset. This involves more than just scraping the web.
- Data Sources: The data includes both bilingual (parallel) and monolingual corpora. Monolingual data is especially crucial for improving fluency and supporting languages with less available parallel text.
- Curation: The process involves significant, sometimes complex, engineering work to efficiently find and match sentences across large domains, ensuring the data is clean and properly distributed across all languages. This curated data is essential for training models to handle nuances that generic web scrapes might miss.
Infrastructure and Engineering at Scale
Engineering is as critical as research at DeepL, with a roughly 50/50 split between the two disciplines.
- Custom Data Centers: DeepL began building its own GPU data centers in 2017 out of necessity, as acquiring sufficient GPU compute was a major challenge. This involved not just racking servers but also building custom frameworks for managing training and inference workloads.
- Inference Optimization: A significant engineering challenge is managing large-scale inference. The system must balance low latency for users with high GPU utilization. This involves dynamically batching user requests and scheduling models based on real-time demand (e.g., spinning up more models for Japanese during its timezone’s peak hours).
- Vendor Lock-in: Currently, the infrastructure is built entirely on NVIDIA GPUs. While alternatives are constantly evaluated, the high cost and engineering effort required for migration, especially with custom model architectures, makes switching difficult.
Quality, Context, and Hallucinations
For modern translation, quality is defined by more than sentence-level accuracy.
- Context is Key: The biggest challenge is providing models with sufficient context. A single sentence translated in isolation can be ambiguous; understanding the document type, the author’s intent, or even company-specific information is necessary for a perfect translation.
- Enterprise Personalization: Fine-tuning a unique model for each of their hundreds of thousands of customers is not scalable. Instead, DeepL enables personalization by allowing models to ingest customer-specific context, such as terminology glossaries, at inference time.
- Hallucinations: While translation models can hallucinate, the problem is less severe than in general-purpose LLMs. The presence of a source text provides a strong constraint, allowing for cross-checking. Specialized models tend to hallucinate less than general-purpose models when used for translation.
The Evolution to Speech and Workflows
- Speech Translation: Speech is a newer frontier where the technology has only recently become practical. The primary challenges are handling the “garbled” and less-structured nature of spoken language and overcoming errors from the initial speech-recognition step. The main engineering focus for speech is achieving near-real-time latency to ensure a fluid user experience.
- Human-in-the-Loop: AI will handle the vast majority of “boilerplate” translation. However, human translators will remain essential for high-stakes, compliance-driven use cases (e.g., legal, financial, life sciences) to provide oversight, check for nuanced errors, and apply deep world knowledge that models still lack. Incremental quality improvements in the AI model provide a high ROI by reducing the time these expensive human experts spend post-editing.
- Future Direction: The future is less about translating sentences from A to B and more about understanding and integrating into the entire enterprise workflow. This involves solving higher-order problems, such as handling document revisions where a human has already corrected a previous AI translation, and feeding that entire context back into the model for a more accurate result.