AI Localization: Automated Translation and i18n Audit

You rolled out the interface in German — the "Save" button went off-screen, dates are in American format, and the error message is still in English. Each manual addition of a new language turns into 3–4 weeks of fixes: errors in plural forms, lost placeholders, inconsistent terminology. We automate this process: from code audit to automatic translation with context awareness. Over many years, we've handled dozens of projects for clients in fintech, e-commerce, and SaaS — average localization time savings reached 60%. For instance, for one fintech product, we cut the localization cycle from 3 months to 2 weeks, saving the company over $50,000 per release.

A typical case: a fintech startup with a React interface in 8 languages. After audit, we found 1200 hardcoded strings, 300 of which broke layout on RTL languages. Implementing i18n + AI translation shortened the release cycle from 2 weeks to 2 days.

Why Internationalization Is the Foundation of Localization

Without proper i18n architecture, any translation breaks layout and logic. Main issues in existing projects:

• Hardcoded strings instead of i18n keys
• String concatenation instead of placeholder formatting
• Ignoring plural forms (in Russian — 4 forms: 1, 2-4, 5+, 0)
• No support for RTL (Arabic, Hebrew)
• Hardcoded date and number formats

# Bad: concatenation
message = "Найдено " + str(count) + " результатов"

# Good: ICU MessageFormat
message = t("search.results_count", count=count)
# In locale file: "search.results_count": "{count, plural, one {Найден # результат} few {Найдено # результата} many {Найдено # результатов} other {Найдено # результата}}"

How AI Analysis of the Codebase Identifies Bottlenecks

We scan the repository using an AST parser and machine learning. The system finds:

• All hardcoded strings (AST analysis + regular expressions)
• Date/number formatting without using the Intl API — MDN recommends this API for localization
• String concatenations with variables
• Images with embedded text (OCR)

class I18nAudit:
    def audit_codebase(self, repo_path: str) -> AuditReport:
        issues = []
        for file in self.scan_files(repo_path, extensions=[".ts", ".tsx", ".jsx", ".py"]):
            ast_tree = parse_ast(file)
            for node in ast_tree.string_literals:
                if not self.is_in_i18n_call(node) and self.looks_like_ui_text(node.value):
                    issues.append(I18nIssue(
                        file=file,
                        line=node.line,
                        text=node.value,
                        issue_type="hardcoded_string",
                        suggested_key=self.suggest_key(node.value)
                    ))
        return AuditReport(issues=issues, summary=self.summarize(issues))

How AI Understands That "Save" Is Both a Button and an Action?

Ordinary machine translation (MT) outputs "Сохранить" for both cases. Our system considers context: element type (button, header, message), screen, user role. A terminology glossary ensures consistency — one term is translated the same way throughout the application.

def translate_with_context(
    key: str,
    source_text: str,
    context: UIContext,
    target_lang: str,
    glossary: Glossary,
    tm: TranslationMemory
) -> Translation:
    tm_match = tm.find_match(source_text, min_similarity=0.85)
    if tm_match and tm_match.similarity > 0.95:
        return tm_match.translation
    terms = glossary.find_terms(source_text, target_lang)
    translation = mt_engine.translate(
        text=source_text,
        target_lang=target_lang,
        context=f"UI element: {context.element_type}, screen: {context.screen_name}",
        enforce_terms=terms
    )
    tm.store(source_text, translation, target_lang, context)
    return translation

According to our data, context-aware translation reduces post-editing corrections by 60% compared to direct MT, further saving the localization budget.

Pseudolocalization: How to Test Localization Before Translation?

Before the real translator starts, we run pseudolocalization: replace characters with decorated ones (e.g., [Ħȇŀŀǿ]) and expand strings by 30% — simulating German or Finnish behavior. This immediately reveals text truncation in UI, incorrect placeholder markup, and hardcoded element sizes.

Continuous Localization: How Not to Break CI/CD?

Integration with TMS (Crowdin, Lokalise, Phrase) via API: with each commit, new strings are automatically sent for translation. QA check before publication: string length, placeholder integrity, absence of machine artifacts. The whole process takes minutes, not days.

Steps for Implementing AI Localization

What's Included in the Work?

1. Codebase audit — identification of all i18n issues (report with recommendations). Takes 2–3 days.
2. i18n infrastructure implementation — framework setup, string formatting. Up to 2 weeks.
3. TMS and glossary setup — connection to Crowdin/Lokalise, terminology creation. 1 week.
4. Translation automation — integration of AI engine with context. From 2 weeks.
5. Pseudolocalization and QA — layout testing before translation, string validation.
6. Release support — monitoring new strings, automatic translation.

Timelines: from 2 weeks (audit + basic implementation) to several months for deep localization of 10+ languages. Cost is calculated individually per project.

Advantages of AI Localization

Over many years, we have implemented dozens of projects in fintech, e-commerce, and SaaS. We guarantee terminology consistency and full placeholder coverage. Automation allows entering new markets 3 times faster compared to manual approach.

Contact us for a consultation — we'll show how localization automation can accelerate your entry into new markets. Request a codebase audit: get a free analysis of one of your repositories.

NLP Development: Text Classification, NER, Embeddings, and Information Extraction

We often receive a task: process 50,000 support tickets — currently all manual. Dataset — 3,000 labeled examples, 12 categories, imbalance: one category occupies 40% of the sample, three at 1-2% each. Baseline accuracy — 78%. Sounds decent until you look at recall for rare classes: 0.31, 0.44, 0.28. These classes — complaints and churn threats — are most important to the business.

This is a typical NLP development project. The problem is not the algorithm but that accuracy is the wrong metric. Our experience across 30+ projects shows: we start by analyzing business metrics and only then choose the model.

Why accuracy is not the right metric for rare classes?

Accuracy ignores imbalance. If the "churn" class appears in 2% of cases, the model can predict "all good" and get 98% accuracy — but the business loses clients. Solution: F1 macro (averaged over all classes) or weighted F1. For NER — strict entity F1 (exact matches only). We guarantee: after choosing the correct metric, model quality becomes measurable and predictable.

Text Classification: From BERT to Distillation

BERT-like models are the standard for classification. ruBERT-base or ruBERT-large from DeepPavlov for Russian. multilingual-e5-large — for multiple languages in one pipeline. XLM-RoBERTa-large — a strong multilingual backbone.

Fine-tuning for classification: add a classification head on top of the [CLS] token, train for 3-5 epochs with lr=2e-5, weight decay=0.01. For imbalance — weighted CrossEntropyLoss or focal loss with gamma=2.0. Contact us — we will show a code snippet.

Imbalance case study. Dataset — 3,000 examples, imbalance 1:20. Solution: class_weight via sklearn + CrossEntropyLoss. Additionally — augmentation of rare classes via backtranslation (ru→en→ru through MarianMT). Recall for rare classes rose from 0.31 to 0.67 with a slight drop in accuracy (76%→74%). Full NLP development end-to-end took 3 weeks.

Distillation for production. BERT-large gives F1 0.89, but inference on CPU — 180ms. Distillation into DistilBERT or ruBERT-tiny2 reduces latency to 25ms with F1 0.84. Export to ONNX Runtime provides an additional 1.5-2x speedup. DistilBERT achieves 7x lower latency than BERT-large with only a 5% drop in macro F1 – a typical production trade-off.

How to choose between BERT and LLM for your task?

For most classification and extraction tasks, BERT-sized models offer the best trade-off between cost and performance. Shift to LLMs only when the task demands generation, complex reasoning, or zero-shot generalization.

NER: Named Entity Recognition

NER — extracting persons, organizations, locations, dates, amounts, document numbers. For general categories (PER, ORG, LOC), pre-trained models work well. For specialized ones (medical terms, legal concepts) — fine-tuning is needed.

Data annotation. The main cost of an NER project. For a quality model — 500-2,000 labeled sentences per entity type. Tools: Label Studio (open source) or Prodigy (by spaCy creators). IOB2 format — standard.

Architecture. Token classification on top of BERT: each token gets a label (B-PER, I-PER, O). spaCy 3.x with transformer pipeline — a convenient production choice.

Nested entities. Standard IOB models cannot handle nested entities (organization inside an address). For such tasks — span-based NER: SpanBERT or SpERT. More complex but correct.

Post-processing is mandatory. The model predicts tokens — normalized entities are needed. Date — dateparser. Amounts — regex + validation. Names — deduplication via rapidfuzz. Included in our standard delivery.

Sentiment Analysis and Opinion Mining

Binary classification positive/negative works out of the box with BERT. Complexity — aspect-based sentiment analysis (ABSA): "the restaurant has good food but terrible service." For ABSA: aspect extraction (NER) + sentiment per aspect. Joint models BERT-for-ABSA — quality on Russian data is lower due to dataset scarcity. RuSentiment, SentiRuEval — main resources.

For production with simple positive/negative/neutral: distil models are enough. Three classes, balanced dataset, 2,000+ examples — F1 macro 0.82-0.87 in 1-2 days.

Text Summarization

Extractive summarization (select sentences) — TextRank or BM25 without training. Fast, no hallucinations. Good for long documents.

Abstractive (generates new text) — seq2seq: mT5, mBART, FRED-T5, ruT5-large. For production via LLM API (GPT-4, Claude) — often the best cost/quality/speed trade-off.

Embeddings: Vector Representations of Text

Embeddings are the foundation of semantic search, deduplication, clustering, RAG. Quality critically affects downstream tasks.

Models. E5-large-v2, BGE-M3, multilingual-e5-large — strong multilingual embedders. sentence-transformers/paraphrase-multilingual-mpnet-base-v2 — fast option. For Russian: ru-en-RoSBERTa (Skoltech) performs well on semantic textual similarity.

Embedding quality evaluation uses the MTEB benchmark as standard. But top results on MTEB don't guarantee success on a domain dataset — we build domain-specific eval.

Fine-tuning embeddings. If standard models don't give the required Recall@k — contrastive learning on domain pairs with MultipleNegativesRankingLoss. How to perform this for domain data:

1. Collect 500–2,000 semantically similar pairs from your domain.
2. Apply MultipleNegativesRankingLoss with a batch size of 32–64.
3. Train for 1–3 epochs using AdamW (lr=2e-5).
4. Evaluate Recall@k on a held-out domain test set.

This approach yields a 5–15% improvement in Recall@k in practice.

Dimensionality and storage. E5-large: 1024 dim, float32 — 4KB per vector. For 10M documents — 40GB. Quantization int8 reduces to 10GB. FAISS IVF_PQ — more compact but with losses. Included in our deployment recommendations.

Information Extraction

Structured extraction is a frequent task. Examples: key contract terms, technical characteristics, dates and amounts from invoices.

1. Regex + rule-based. For INN, OGRN, amounts, dates — more reliable than neural networks. No data required.
2. NER + post-processing. For variable formats.
3. LLM with structured output. GPT‑4 / Claude with JSON schema — for complex documents. Cost: minimal per document. For 10k+ documents/day — we calculate the economics.

We guarantee a hybrid: regex/NER for typical fields + LLM for edge cases. Our guarantee is backed by years of production experience and more than 30 projects.

Work Stages

Prototype on existing data — 1-3 weeks. Production system with CI/CD — 1.5–2.5 months. Cost is calculated individually — get a consultation for a project estimate.

What's Included

• Model and pipeline architecture documentation
• Access to the model via REST API (FastAPI + ONNX)
• Client team training (2-hour webinar + Q&A)
• Accuracy guarantee on the agreed test set
• Months of post-delivery support (bug fixes, adaptation to new data)

Our Experience

Years of NLP projects from classification to RAG systems. The team includes ML engineers experienced with Hugging Face, spaCy, LangChain, MLOps. We use vLLM, Kubeflow, Weights & Biases — a production stack, not toys. Contact us to evaluate your NLP project within two days — request a free consultation on your text processing pipeline.