Why AI Fails on Geological Data, and What Fixes It | Full Webinar

Source-verified mining data for geology. The full Pulse Intelligence webinar, 17 September 2026, on digitising and normalising geological records so AI can work on them at scale: WAMEX reports, NI 43-101, JORC and SK-1300 technical reports, drill logs, and 120 year old Cyrillic archives. Point a large language model at a folder of raw geological PDFs and it hands you a confident answer with very little behind it. Ask a general model for every ASX company under half a billion dollars market cap ranked by enterprise value per ounce and it returns five or six names when the real answer is closer to 120. The fix is not a better prompt. It is the condition of the data underneath it: units of measure across competing reporting standards, geospatial data that thins out the older the record gets, and a century of evolving interpretation, mathematics and coordinate systems. Will Coetzer and Vladislav Tiazhov of Pulse Intelligence show how that is built. Chris Yeomans of Neo Geo Consulting, an independent PhD geologist specialising in geodata and AI due diligence, spends the hour arguing with them. CHAPTERS 0:00 Welcome and what this session covers 6:29 What breaks a geological record 11:58 Why general AI gets mining questions wrong 14:00 Every answer you get from an AI is a hallucination 19:18 The spatial problem: terabytes you cannot search 20:07 The platform: tenure, WAMEX, drill holes, geochemistry 30:47 Kazakhstan: three terabytes of Cyrillic archives 37:16 Normalising metre by metre drill logs to modern intervals 39:15 Geo inference: aligning historical maps by their railways 46:20 Where AI should stop and the geologist starts 55:34 Ranking the credibility of a source HOW ACCURATE IS THE DIGITISATION? The Kazakh archive is three terabytes of scans, up to 120 years old, mostly Cyrillic, none of it digitised. OCR: about 99% character accuracy on low quality scans, which makes the archive searchable in Russian and in English. Drill passport extraction: 98.9% accuracy against a ground truth set transcribed by hand. Lithology, age, intervals and multi-element assays. Georeferencing: 30 of 30 test maps placed correctly, at about 2% drift against a manual ground truth. Extraction alone is not enough. Soviet era logs are recorded metre by metre, so a query for every intercept over five metres grading better than 2% copper runs through millions of rows. Normalised to modern interval reporting, it is one question. HOW DO YOU GEOREFERENCE A HAND DRAWN MAP? In two passes. First, digitise the map, read the settlements off it and match those names to their modern equivalents, allowing for towns that were renamed. That left three to five percent drift. The second pass aligns the map by its railways, matching the curve of the line to a global railway layer. HOW DOES PULSE DECIDE WHICH SOURCE TO TRUST? An assay certificate, a company announcement and management commentary are not equal evidence. Each document type is categorised separately at extraction, then ranked for credibility with technical reports at the top, and figures are cross-referenced against other credible sources. Every value is timestamped and opens its source page in one click. WHERE SHOULD AI STOP AND THE GEOLOGIST START? The argument made here is that AI should compress the search and leave the interpretation alone: give the geologist the questions, not the answers. Chris Yeomans puts the counter case. WHAT IS WAMEX? Western Australia's open-file mineral exploration archive, the reports licence holders must submit to the state. Roughly 15 million pages of maps, drill logs, assays and interpretations, public and close to unusable at scale because it is unstructured paper. Pulse has structured it. WHAT IS THE PULSE MCP? MCP is the Model Context Protocol, the open standard connecting an AI model directly to a data source. The Pulse MCP lets Claude, ChatGPT or your own internal model query Pulse inside your own workflow, restricted to Pulse data, every answer traced to its source filing. ABOUT PULSE INTELLIGENCE The technology and infrastructure layer for mining data. We digitise, validate and structure NI 43-101, JORC and SK-1300 technical reports, exchange disclosure, government archives and historical geological records, then serve them through a platform, a documented API and the Pulse MCP. More than 450 metrics across 29,600+ mining assets. Request access: https://pulseintelligence.com API documentation: https://pulseintelligence.com/api Contact: hello@pulseintelligence.com TOPICS COVERED Geological data digitisation, mining data normalisation, AI in mining and exploration, OCR of historical geological reports, georeferencing historical maps, WAMEX search, NI 43-101, JORC and SK-1300 data extraction API. Less searching. More strategising. #miningai #miningdata #exploration #criticalminerals #geology