06
e-translator
guarded translation routing, edit planning, and validation for technical pdfs
headless Python pipeline, no UI to capture; illustration pending
e-translator is a deterministic PDF translation pipeline. It parses, routes, translates, applies edits, validates the result, and reports explicit product states instead of hiding quality problems.
- supported_with_degradation
- validation artifacts
- PyMuPDF visual backend
Architecture
Nicolás built it on one idea: translating a PDF is orchestrating a document, not one model call. He split parsing and routing, OCR, guarded translation, edit planning, visual application and validation into separate layers, each with one job.
Document graph
The hard engineering layer is the canonical document graph: Nicolás normalizes pages, blocks, spans and layout metadata into one internal representation, and his runtime routes each page between native extraction, Docling and OCR before the guarded translation and the edit planning.
Release gating
Nicolás built it to be honest: when the layout comes out weak, e-translator does not fake success; it says so with explicit states, supported, supported with degradation or unsupported, backed by validation reports, support matrices and visual scorecards.
Runtime contract
It runs as a product, not a pile of scripts: one official command, an older entry kept working for compatibility, and one configuration file that holds the client-facing defaults.