Case study · 2025 · AI/ML · Data · Full-Stack
Product Carbon Footprint Engine
Fine-tuned LLMs read a manufacturer's raw parts list, match every material against a 1M+ record emissions library and produce the product's carbon footprint report — automatically.
98.7%
Parts-list parsing accuracy
1M+
Materials searchable
40%
Lower retrieval latency
00 How it works
A manufacturer uploads a messy, multi-page parts list
01 The problem
To report a product's carbon footprint, every component in its parts list (bill of materials) has to be matched to an emission factor. Manufacturers send these lists as messy, multi-page spreadsheets and PDFs, and matching thousands of rows by hand against a library of 1M+ materials doesn't scale.
02 Approach
- 01
Asynchronous pipelines with local LLMs — fine-tuned with LoRA/QLoRA for parts-list parsing and material classification — turn unstructured, multi-page documents into structured materials.
- 02
Hybrid Qdrant vector search with metadata filtering maps each material to LCA libraries.
- 03
Fuzzy matching plus multi-LLM integration produce a confidence score for every categorisation.
- 04
Token-budget throttlers and NDJSON + SSE streaming keep LLM inference fast and predictable.
03 Architecture
04 Results
- 98.7% accuracy parsing unstructured, multi-page parts lists.
- Semantic matching across a library of 1M+ materials.
- 40% lower retrieval latency through token budgeting and streaming.
- Automated a manual parts-list-to-emissions reporting workflow.
Next case study
AnswerGate — Regression Tests for AI Assistants