MiningCave AI Customer Agent
82% of support queries resolved without human intervention
82%
Support Deflection
1,200+
Products Indexed
< 2 seconds
Response Time
24/7
Availability
The Problem
MiningCave is an Australian mining equipment retailer with a catalog of 1,200+ highly technical products (drills, pumps, safety gear, underground machinery), each with compatibility charts, operating specs, and compliance requirements. Their support team was spending most of their time answering pre-sale questions that were already documented in the product pages and manuals. They needed a way to handle more volume without hiring more people.
Our Approach
Three stages. A scraping layer that pulls all product data (specifications, compatibility charts, PDF manuals, FAQs, and category metadata) and keeps it in sync with the live catalog daily. An ML layer that chunks, embeds, and indexes all of it into a vector database with hybrid search tuned for technical specification queries. An AI agent deployed as a website widget that answers natural language questions using retrieved catalog context, generates grounded responses, and escalates to human support when confidence falls below threshold.
Pipeline Breakdown
01 · Collect
- Full catalog scraping: 1,200+ products, specs, and manuals
- Daily sync to capture catalog updates and new product listings
- PDF manual extraction, text normalization, and structured storage
02 · Process
- Content chunking strategy optimized for technical Q&A
- Dual-encoder embeddings for semantic and keyword retrieval
- Hybrid search index combining dense vectors and BM25
- Confidence scoring for human escalation routing
03 · Act
- Real-time query answering via embedded website chat widget
- Escalation to human support when confidence < 0.7
- Feedback loop: low-confidence queries flagged for review
- Weekly analytics report on top questions and deflection rate
Where similar patterns could help
These links describe possible applications of related engineering patterns. They do not mean this project was delivered for those industries, and its results are not a forecast for another business.
Distributor RFQ workflows
Explore catalog normalization and product matching for a proposed RFQ-to-quote workflow.
Explore the workflowDistributor order-entry workflows
MiningCave's catalog collection and structured product information provide transferable engineering context for order-entry preparation, not evidence of a distributor order-entry deployment.
Explore the workflowDistributor product-data preparation
MiningCave’s catalog collection, normalization and structured storage provide transferable engineering evidence for reviewed supplier imports, not a distributor master-data or PIM deployment.
Explore the workflowHave a similar operational bottleneck? Discuss your project
← Back to all work