You’ll own how accurate Volta's AI is, and you’ll make it measurably better, customer by customer. You’ll build the evaluation harnesses, run the error analysis and improve extraction and matching on real purchase orders from enterprise customers. You’ll report to the Head of AI: this role is about measurement, retrieval and LLM-based extraction, not about training models.

- Own the accuracy number: we are working towards 90% accuracy across customers. Define how it's measured per field (customer, ship-to, product, quantity, unit of measure, price), track it per customer and per release, and make it go up

- Build evaluation that runs itself: per-customer evaluation harnesses and golden datasets built from real, anonymised orders, wired into CI so a regression is caught before release, not by the customer

- Make matching better: product and customer matching against catalogues of thousands of SKUs and messy master data, with Elasticsearch-based retrieval, synonyms, reranking and confidence scores

- Improve extraction with LLMs: prompts, schemas and pre-processing steps that turn an email body table, a multi-page PDF or a 100-line spreadsheet into clean order lines

- Do error analysis like an engineer: turn a pile of customer issues into categories, root causes and the one fix that closes ten tickets, then verify in production after the release

- Turn customer fixes into configuration: per-customer behaviour lives in prompts and configuration, not in code branches, and recurring patterns become reusable templates by industry and document type

Applied data scientist who measures things:



3+ years shipping ML or NLP to production, judged on a metric that mattered to users LLM applications in production: extraction, classification or retrieval built on LLM APIs with structured outputs and real evaluation, not only demos or notebooks Search and matching: entity resolution, record linkage or product search, ideally with Elasticsearch, OpenSearch or hybrid lexical plus vector retrieval Evaluation discipline: test sets, metrics, label noise and regression tracking over time Real-world documents: messy PDFs, spreadsheet exports, emails in several languages Writes production code: Python that other people can review, test and run in CI Must have: strong Python, pandas or polars and testing habits; hands-on with LLM APIs (Gemini, OpenAI, Anthropic or similar); evaluation datasets and metrics (precision and recall, per-field accuracy, confidence calibration); search and retrieval (Elasticsearch or similar, fuzzy matching, BM25, embeddings, reranking); solid SQL; rigorous error analysis; fluent English; French or Italian (many documents are written in them) Nice to have: LLM tracing and evaluation tooling (Langfuse or similar); document parsing and OCR; B2B commerce data (catalogues, units of measure, price lists, ERP master data); GCP (Cloud Run, Vertex AI) Degree: MSc in Computer Science, Statistics, Applied Math, Computational Linguistics or a related quantitative field, or equivalent practical experience Location: Milan or Paris, hybrid At Volta, we believe diversity makes our teams stronger. We welcome all applications, regardless of gender, background, or experience.

📌 Data Scientist, AI (Lombardia)
🏢 Volta Software
📍 Lombardia