IP engineers were manually scanning and tagging multilingual patents - slow, inconsistent, and hard to replicate at scale. Multilingual content and unstructured abstracts made it difficult to compare filings and report consolidated trends to innovation stakeholders.

Built REST endpoints for language classification → translation → domain-specific embedding → unsupervised clustering → cluster summarization → trend tracking, exposed via a FastAPI service on Azure Databricks and backed by a vector index for semantic lookups. Clusters receive concise titles/summaries and can optionally align to IPC categories for consistent reporting.

Baseline EV-battery patent trend analytics high-level solution architecture

• Cloud: Microsoft Azure Cloud Infrastructure
• CI/CD: Azure DevOps Pipelines
• Containerization: Docker
• Data Platform: Azure Databricks (APIs for ingestion & jobs)
• Vector Index: Databricks Vector Search
• Backend: Python services with FastAPI (REST)
• Language Detection: XLM-roberta
• Translation: mBART-large-50 many to one
• Embeddings: BERT-for-patents (fine-tuned for EV-battery patents)
• LLM (summarization): Mistral 7B

• Multilingual filings: Used XLM-RoBERTa + mBART-50 to standardize language before embedding.
• Noisy abstracts & jargon: Fine-tuned domain embeddings (BERT for Patents) to boost semantic cohesion before clustering.
• Inconsistent labels: Auto-titled clusters with Mistral-7B; optionally mapped to IPC-labels.
• Scalability & repeatability: Orchestrated Databricks jobs with containerized services and Azure DevOps CI/CD for stable weekly runs.
• Analyst adoption: Added concise summaries + trends dashboard, shifting effort from manual tagging to interpretation.

• Earlier visibility of emerging battery-tech themes; analysts redeployed to higher-value analysis.
• Analysts reallocated from manual tagging to higher-value trend interpretation
• -90% review/labeling time per weekly batch (from 5 days → 0.5 day)
• ≈ €490k/year cost savings (based on 10 IP engineers' time saved)