Michele Marchi

• Deployed production FastAPI services for invoice-labeling and RAG, backed by automated tests in the deployment workflow. • Built an LLM-as-a-judge framework benchmarking 8 Qwen variants, driving a 30% accuracy gain for production model selection. • Delivered 30 Langflow agents from experimentation to production, documenting agents and components for client demos. • Shipped an OCR pipeline (PaddleOCR, Qwen-VL) processing 2,000 contracts at 100 documents/hour, demonstrated live to the client.

• Produced Plotly dashboards and reports, presenting monthly progress to non-technical stakeholders. • Implemented end-to-end RAG pipelines in Python, combining BERT retrieval and Llama generation on local HPC. • Built end-to-end data pipelines (ingestion, cleaning, review) for LLM fine-tuning, integrating ~200 curated entries. • Developed a Streamlit app for human-in-the-loop annotation, replacing a spreadsheet-based curation workflow.

• Managed corpus version control in Google Sheets with dual-column tracking while cleaning hundreds of dialogue entries. • Built Python scraping pipelines with BeautifulSoup, extracting conversation data from ~12 websites into TSV for training. • Fine-tuned Llama and Pythia with PyTorch, evaluating outputs through conditional perplexity and HTER. • Co-authored a research paper accepted at ACL, covering experimental design and results analysis.