Lettria

Lettria provides a visual workflow editor that chains pre-built NLP components (tokenization, entity extraction, sentiment analysis, classification) without requiring code. Users drag components onto a canvas, configure parameters through UI forms, and the platform generates the underlying processing graph that executes sequentially or in parallel. The builder abstracts away model selection, hyperparameter tuning, and deployment complexity by exposing only business-relevant configuration options.

Lettria's entity extraction engine uses pre-trained language models that support 40+ languages out-of-the-box, enabling users to extract entities (persons, organizations, locations, products) from text in multiple languages without retraining or language-specific configuration. The system likely leverages transformer-based models (e.g., multilingual BERT or XLM-RoBERTa) fine-tuned on diverse language corpora, with a unified inference pipeline that handles language detection and entity boundary detection across scripts and morphologies.

Lettria provides sentiment analysis that classifies text into polarity categories (positive, negative, neutral) and optionally detects emotions (joy, anger, fear, surprise). The implementation uses pre-trained classification models (likely fine-tuned transformers) that score text against learned sentiment patterns. Users can configure the granularity of sentiment output (binary positive/negative vs. multi-class) and set confidence thresholds through the UI, with results returned as structured scores and labels.

Lettria enables users to define custom text classification categories (e.g., 'product inquiry', 'complaint', 'feature request') and train classification models by providing labeled examples through the UI. The platform uses active learning or semi-supervised learning patterns to minimize the number of labeled examples required, likely leveraging transfer learning from pre-trained language models. Users upload labeled training data (CSV or JSON), the platform trains a classifier, and returns a model that can be deployed via API or used in pipelines.

Lettria exposes all NLP capabilities through a REST API with standard HTTP methods, allowing developers to integrate text processing into applications, microservices, and workflows. The API accepts JSON payloads with text and pipeline configuration, returns structured JSON responses with results, and supports batch processing for high-volume use cases. Webhook support enables asynchronous processing and event-driven architectures, where Lettria sends results back to a specified URL when processing completes.

Lettria supports bulk processing of text data through CSV and JSON file uploads, allowing users to process hundreds or thousands of documents in a single batch job. Users upload files with text columns, specify which NLP pipeline to apply, and receive results as downloadable CSV or JSON exports. The platform handles file parsing, applies the pipeline to each row, and aggregates results with metadata (processing time, error logs) for quality assurance.

Lettria allows users to save, version, and deploy NLP pipelines as reusable components. Users can create multiple versions of a pipeline (e.g., 'sentiment-v1', 'sentiment-v2'), compare versions, and promote specific versions to production. The platform manages deployment endpoints, tracks which version is active, and enables rollback to previous versions if new versions underperform.

Lettria provides dashboards and reports showing pipeline performance metrics such as processing latency, throughput, error rates, and result quality indicators. Users can view execution logs, sample results, and confidence scores for each pipeline run. The platform may track metrics like entity extraction precision/recall (if ground truth is provided) or classification accuracy on labeled test sets.

Lettria enables users to chain multiple NLP components into complex workflows with conditional branching. For example, a pipeline might first classify text into categories, then apply different entity extraction rules based on the category, or route text to different sentiment analysis models based on language detection. The platform provides if-then-else logic nodes and supports sequential and parallel execution of components.