Rising Innovators
Nordlys – GenAI-Powered Supplier Intelligence for Industrial Procurement
Nordlys applies domain-specific Matryoshka embeddings and LLM reranking to semantically match industrial Bills of Materials against 2M+ suppliers, replacing manual sourcing with explainable AI at scale.
B2B SaaS platform applying generative AI to industrial supply chain procurement. Incubated at I3P – Politecnico di Torino, we match complex Bills of Materials to a network of 2M+ suppliers using domain-specific embeddings and LLM reranking.
Motivation
We applied to GenAI Zürich because we believe the most impactful generative AI won't replace what we Google — it will transform the industries that still run on phone calls and spreadsheets. Industrial procurement is a $12T market with near-zero AI penetration. We built Nordlys to demonstrate that frontier embedding architectures, applied to unglamorous but structurally complex B2B verticals, can deliver measurable, production-grade results. GenAI Zürich is the right stage to show that domain-specific AI is not a niche, it is the next frontier.
Lapo Chirici
Chief Information Officer
Industrial procurement in Oil & Gas, Energy, Aerospace, and Advanced Manufacturing still runs on manual scouting and static supplier databases. Over 65% of supply chains remain locked with the same vendors for decades. General-purpose AI fails here: matching a Bill of Materials requires parsing metallurgical specs, certifications, process capabilities, and business constraints — not keywords.
Nordlys is a B2B SaaS platform built on a proprietary multimodal embedding pipeline. Each supplier is represented by at least four independent semantic vectors (services, products, industry context, long-form description), trained using Matryoshka Representation Learning — enabling adaptive dimensionality from 64 to 768 dimensions for speed-accuracy balance. When a buyer uploads a BOM (Excel, PDF, or CAD/CAM files), our engine projects it into the same vector space and computes a weighted composite similarity score across all four fields. A final LLM reranking layer (R_LLM, scored 0–3 by Gemini for task-result coherence) multiplies the vector score to produce a unique, non-tied ranking — with every decision traceable. Business filters (geographic, revenues from Chamber of Commerce data, lead time, financial health, references, ESG compliance) gate the final output.
Our system achieves 70%+ matching accuracy versus a 25% baseline with keyword tools — a +45 percentage point delta validated through a proprietary multi-LLM evaluation framework. In 2025 we completed 11 paid proof-of-concepts generating €175K in revenue, including a strategic pilot with a global OEM (772 components, T54 product line). All data processing is GDPR-compliant, hosted on EU infrastructure, with planned migration to Switzerland. A proprietary industrial LLM is in development for Q3-Q4 2026 release.