Gina Welsh

I build production RAG systems and evaluate LLM output quality using Python, PostgreSQL, Azure, Hugging Face. I am a computational linguist by background, with native-level English and fluency in French and Swedish. Demonstrated experience with speech processing (TTS and ASR fine-tuning). I can handle the hard parts that many AI engineers skip: evaluation, AI-text detection, and multilingual/low-resource NLP.

EU-based with Irish (EU) citizenship.

Showcased projects:

RAG system for financial document analysis

Designed and built a production retrieval-augmented generation system that lets non-technical users query a large body of IT financial documents in plain language and get answers grounded in the source material — with citations back to the original documents, so results are auditable rather than a black box. Full stack: retrieval layer, LLM answer generation, PostgreSQL backend on Azure. Ongoing work includes performance tuning and reliability improvements as usage scales.

Multilingual speech recognition for a low-resource language

Fine-tuned a Whisper speech-to-text model to work on a language with very little available training data, by generating a synthetic speech dataset and combining it with real dialectal recordings. Ran systematic experiments to find the best mix of synthetic vs. real data, and measured accuracy gains with standard error metrics (WER/CER). Relevant for any client needing speech or NLP capability in a language the big models handle poorly.

Bias evaluation and mitigation for machine translation

Evaluated and reduced gender bias in translation systems, testing two mitigation methods across two production translation models on a purpose-built test set. Delivered a clear finding on which method worked best and a framework that generalises to other languages. Peer-reviewed and accepted at a European NLP conference (EAMT 2026).