80%
Manual work reduction
Case Study
Reducing Manual Work by 80%: AI Data Classification in Mining
Days → Minutes
Classification time reduced
94%
Classification accuracy
3 engineers
Built the full platform
Full-stack Platform Build ● Production AI Data Engineering
Most teams trying production AI focus on the model. That’s the easy part. The hard part is everything around it: the workflow that fits how people actually use the tool, the human review layer that catches what the AI gets wrong, the validation that makes the output trustworthy enough to ship.
This shows up clearly in AI data classification. The model handles one language fine and chokes on the next. Accuracy looks great on a clean test set and collapses on real spreadsheets where the same equipment failure shows up under five different names. Domain experts won’t trust the output, so they re-check every row, and the automation saves no time at all. The model isn’t where these projects fail. They fail in everything around the model: language handling, taxonomy enforcement, review workflows, the architecture that decides when rules should override predictions.
This shows up clearly in AI data classification. The model handles one language fine and chokes on the next. Accuracy looks great on a clean test set and collapses on real spreadsheets where the same equipment failure shows up under five different names. Domain experts won’t trust the output, so they re-check every row, and the automation saves no time at all. The model isn’t where these projects fail. They fail in everything around the model: language handling, taxonomy enforcement, review workflows, the architecture that decides when rules should override predictions.
Manual Classification was the bottleneck
The work that has to happen before the work
Volatile commodity prices and shifting market valuations are pushing mine operators to cut costs, limit capital expenditure, and squeeze more from existing operations. But the size and speed of profitability improvements vary wildly by commodity, region, and individual mine. Our client, a top-three global management consultancy, built a platform to close that gap: analyzing labor, cost, and equipment-productivity data, benchmarking it against a global peer set, and surfacing where the real performance-improvement opportunities are.
The issue was what happened before any of that analysis could begin.
Every mining company records things differently. A truck breakdown might show up as “Mech Engine” in one file and “Avería mecánica del motor” in another. Different formats, different languages, different terminology. Before any real analysis could start, a consultant had to open each spreadsheet and classify every row into a standard taxonomy. Thousands of rows. By hand.
This happened on every engagement. Consultants were spending days on classification before they could get to the analytical work their clients were actually paying for.
A trustworthy AI solution, not just accurate
Accuracy is table stakes. Trust is the bar.
The client wanted to automate classification. But accuracy matters when the output feeds directly into client-facing analysis. A model that gets it right 60% of the time creates more cleanup work than it saves.
They needed a production platform: an AI that classifies across languages, a human review layer so domain experts stay in control, and an interface that fits into how consultants actually work. Upload, map, classify, review, export. No room for missed steps or inconsistent process.
The platform, from scratch
Classification, review, and analytics in one workflow
Together with the client, we built an 8-step guided workflow that takes raw mine site spreadsheets and produces analysis-ready datasets. Upload a file, map the relevant columns, and the AI classifies everything in minutes, reducing manual classification work by approximately 80%.
The taxonomy covers real operational depth: 129 cost categories, 33 labor roles, and roughly 160 equipment delay types across four equipment classes. The platform handles data in any language, translating before classification and preserving original values for export.
Every AI classification goes through human-in-the-loop review. Consultants see each suggestion, accept or override it, and export a clean dataset. Nothing is finalized without a human decision.
We also built an analytics engine on top. From classified equipment data, the system generates 13 pre-formatted analysis tables (OEE breakdowns, maintenance reliability, utilization summaries, benchmark comparisons) packaged as a single download.
The AI accuracy story is worth telling. Equipment delay classification started at 56% and was improved to 94% at the category level through prompt engineering, targeted examples, and a hybrid architecture: AI handles granular prediction while a rules engine enforces category-level consistency. During evaluation, the team found 79 cases where the AI was right and the human-labeled training data was wrong.
What changed
Manual classifaction work reduced by 80%, 94% accuracy, days to minutes, one platform.
Classification that used to take days now takes minutes. Consultants moved from data entry to quality review. Results are consistent across engagements regardless of source language or site format.
The client now has a platform that handles files up to 500 MB, covers three active classification domains (with a fourth in development), and runs on infrastructure built for long-term use: .NET 9 backend, Azure cloud, 277 automated tests, and performance optimizations that reduced AI API calls by 180x through smart deduplication.
A small team of roughly 5 engineers built the entire production-grade platform from scratch over 9.5 months (341 commits), cutting manual classification work by 80%: AI classification engine, full-stack web application, cloud infrastructure as code, and a rigorous evaluation framework, cutting
Why this matters – a solution beyond mining
AI data classification isn't a mining problem. It's a production AI problem/
Every organization that collects operational data across multiple sites or regions faces some version of this problem. Inconsistent formats, different languages, manual standardization that eats time and introduces errors. The bottleneck is always the same: getting from raw data to something you can actually analyze.
The AI classification model was one piece. The production platform around it (the workflow, the validation, the human review, the multilingual support, the analytics engine) is what made it usable at scale. And that’s the part we built.
Tecknoworks is a Production AI Systems Engineering company. 25 years building software, 15 years in data, 10 years in AI. We build the production infrastructure that makes AI work in the real world.
Our Success Stories
Discover client success stories and use cases, covering extensive topics, including next-gen technologies, data analytics, automation processes, and more.
AI Delivery Squads
Need ongoing capacity? Our engineers embed in your organization and ship production AI continuously. Not a consulting rotation. A dedicated team aligned to your mission, operating inside your systems.