AILY LABS
Data Scientist, July 2023 to Nov 2025, Madrid. I joined straight after my master's and built production AI for a Fortune 500 pharmaceutical company: LLM agents, retrieval, NLP, and time-series forecasting. I started as Junior Data Scientist I and left as Mid Data Scientist.
The company
AILY LABS builds a decision intelligence platform for large enterprises: AI agents that turn company data into recommendations for finance, supply chain, R&D, and commercial teams. It was founded in 2020 by Bianca Anghelina.
- Aug 2023. €19M Series A, about $20M, led by Insight Partners. It was the company's first outside funding, a month after I joined.
- Nov 2025. $80M Series B led by FPV Ventures, with Insight Partners and J.P. Morgan. It was announced the month I left.
Shop-floor agent
I built a Graph RAG agent over factory equipment manuals. An operator enters a fault or alarm code and gets a likely procedure, plus the page and section of the manual to check it against. A confident answer without a page reference counted as a failure.
The ingest step ran OCR on scanned PDFs and kept tables, figures, page numbers, and section names. A graph database linked equipment, faults, and procedures, and a retrieval layer on top returned cited passages, so other agents and applications could use the answer and still show where it came from.
Graph RAG pipeline
Scanned factory equipment manuals, with OCR for diagram pages.
The operator gets the fault, the manual location, and the exact page.
NLP before GenAI
Before generative AI was practical in production, I worked on text classification for quality assurance on factory machines. We compared CatBoost with XGBoost, LightGBM, and deep neural networks. CatBoost won: it trained on GPU, it was the fastest to train, and it gave the best results.
Agents in production
The agents used LangChain for retrieval and tool calls, with a Langfuse trace for every run. When an answer looked wrong, we could see which documents the agent had retrieved, which tool it had called, and where a guardrail had stopped it.
Data pipelines
I owned Airflow and Docker pipelines on AWS. One internal ETL took about 90 minutes. After I restructured the logic and added the right indexes, it ran in 15 seconds.
Team
I mentored interns through data preparation, modeling, evaluation, and deployment, and started a recurring knowledge-sharing session for the team.