Sluicebox Scales AI Agents for Dynamic Life Cycle Assessments With NVIDIA Nemotron

The idea for Sluicebox grew out of a problem that co-founder and CEO Elmar Kert encountered while working in the industry: Teams needed to understand how electronics supply chains fit together, where carbon hotspots were concentrated, and which product or supplier changes could deliver the greatest impact.

The difficulty starts with the data behind a product carbon footprint. Analysts must manually sort through tens of thousands of bills of material records for each product and determine what each component is, what it contains, how and where it was manufactured, how it moved through the supply chain, and which emission factor applies. Much of that information sits outside enterprise systems in multimodal datasets composed of bills of materials, PDFs, CAD files, images, supplier emails, and factory records. Analysts then have to reconcile inconsistent terminology and search emission-factor inventories containing more than 10,000 entries. This entire process can take years to collect the needed primary supplier data.  

Supplier information creates another bottleneck. Specifications arrive in different formats and with gaps that require repeated follow-up. Even when a company completes an assessment, the result can arrive too late to influence the product. By the time many traditional LCAs are complete, the products they cover often have already been on the market for years, turning carbon accounting into a backward-looking reporting exercise.

At the same time, demand for credible product-level data is growing. Large buyers and hyperscalers need information from their upstream suppliers to measure progress against sustainability commitments and pressure suppliers to meet emissions reduction targets. Manufacturers therefore need a way to scale carbon analysis without increasing manual data work. More importantly, they need results early enough to guide the decisions that determine a product's footprint pre-production. 

To tackle this data pipeline issue at the necessary speed and scale, Sluicebox built its AI agents to create a structured model of a product and its supply chain. The agent connects components, materials, suppliers, manufacturing processes, and transportation data in a carbon graph that teams can query and update as new information arrives.

Sluicebox's Dynamic LCA intelligence begins by ingesting multimodal records and converting them into consistent product data. It enriches incomplete bills of materials with information such as component weight, material composition, and manufacturing activity. The resulting model gives teams a clearer view of what a product contains, where its emissions originate, and where data gaps require additional supplier information.  

The company rebuilt three of the most labor-intensive steps as AI workflows:

• Component classification: sorts each bill-of-materials line into the appropriate category in an approximately 100-node taxonomy
• Supplier part-data extraction: turns unstructured emails and documents into clean records that can support an assessment
• Emission-factor matching: maps each component or process to the best available factor in an inventory containing more than 10,000 entries

The team also developed Lucy, its supplier AI agent, to collect and validate primary data. Lucy guides suppliers through information requests, identifies missing fields, and checks whether values fall outside expected ranges. This helps Sluicebox replace broad industry averages with data grounded in how a component was actually produced, and companies can get primary data coverage of 90%+ of their suppliers in weeks instead of years.  

Sluicebox evaluated NVIDIA Nemotron 3 Ultra for these workflows. NVIDIA Nemotron is a family of open models designed for efficient agentic reasoning and fast task completion. In Sluicebox's tests, Nemotron 3 Ultra processed the context needed to classify parts, structure supplier records, and select emission factors while delivering 51–80% lower inference costs, improved accuracy, and response times equal to or faster than the company's prior production model.

Supplier-data collection is a conversation, not a one-time calculation. This means a delayed response can interrupt that exchange and add friction to an already demanding process. Faster inference lets Lucy absorb new information, maintain context, flag possible outliers, and guide a supplier toward a complete response while the interaction still has momentum.