Project case study · Applied NLP · Windows Desktop · Psychology
PsyLex Studio: a local-first psychology translation workbench
An installable bilingual psychology desktop workbench: search local terminology, connect an optional language model, extract scanned documents, review Persian results, and export work—all with a clear boundary between the open-source app and privately imported dictionaries.
01 · Context
Research question
General-purpose translation can miss domain-specific psychology vocabulary, distort statistical statements, or ignore whether a phrase is a verified term or generated text. This project focuses on a usable English-to-Iranian-Persian workflow that keeps local reference evidence distinct from LLM output and offers a practical Windows release.
02 · Methods
Approach
- Developed a Python/PySide6 desktop UI with Dictionary, Translation, Documents, Batch, History, Favorites, and Dictionary & Data workspaces.
- Implemented SQLite-backed glossary search and DOCX terminology imports, keeping user-imported data local and avoiding redistribution of third-party reference books.
- Connected optional local inference servers (Ollama, LM Studio, llama.cpp, vLLM) and supported online providers, with first-launch discovery and Windows Credential Manager-backed API keys.
- Added RapidOCR-based recognition, optional OCR provider integrations, segment-by-segment editing, export, and review signals for psychology terminology.
- Built Windows packaging using PyInstaller and Inno Setup, with Python regression tests, GUI smoke tests, and a packaged executable self-test.
03 · Evaluation
Reported results
- PsyLex Studio v0.6.5 source code and a downloadable Windows installer are published on GitHub Releases.
- The available OCR catalog includes two local options and six online adapters; local model discovery supports four running server families.
- The public-source edition ships a 16-term original starter glossary and can import user-authorized bilingual DOCX terminology.
04 · Scope
Limitations
This is applied desktop software, not a validated clinical translation system or a benchmark showing improved translation accuracy. The sample glossary is intentionally limited; local models require separate runtimes, cloud services may charge, and OCR/LLM output needs expert review. Dependency redistribution licenses are handled separately from the MIT license for original code.