Document understanding
OCR and multimodal language models that read typewritten and handwritten 20th-century logs — including tables that run across page breaks. Every extracted value keeps a link to the page it came from.
Software · AI · Geodata
Now the machine is teaching us what it means.
BitBlocks 01 AB builds systems that turn scanned drill logs, mine plans and assay lists into structured, searchable, visualisable data — and 3D views you can actually walk down into.
What we build
We work where document understanding, geodata and 3D meet — a niche too narrow for the big platforms and too hard for a spreadsheet.
OCR and multimodal language models that read typewritten and handwritten 20th-century logs — including tables that run across page breaks. Every extracted value keeps a link to the page it came from.
Python pipelines on Polars and DuckDB, columnar storage for large tables, and search that combines semantics, a knowledge graph and exact text matching. Many owners in one system, hard isolation between them.
Drill holes, terrain, mine plans and block models in an ordinary browser — no installs, no special hardware. The same scenes can be stepped into with a VR headset.
Platform
Skarn is the platform everything else is built on. It takes geological documents, drill data and field observations and turns them into a knowledge base you can question — cited down to page and interval.
It is under active development and runs in earnest on our own projects before anything is offered to anyone else. Get in touch if you want to know where it stands today.
Data sources
SGU's open geodata is the floor, not the ceiling. Skarn also reads historical mine archives, public registers, licensed datasets and the client's own drill and field data — and keeps track of what each of them may be used for.
Bedrock, geophysics, drill holes, mineral rights and geochemistry. The baseline everyone has — fetched and kept current automatically.
Scanned reports, mine plans and drill logs from the twentieth century, often handwritten. What exists but never became machine-readable.
Property, permits, protected areas and infrastructure. What decides whether a deposit can be worked in practice, not only in the rock.
Commercial geodata, price series and industry data that require an agreement. The terms travel into the system instead of staying in the contract folder.
Drill core, assays, field observations and earlier studies. Often the most valuable material — and the one that must never leak.
Four sources may be combined and merge into one model. The licensed one stops at the gate: still searchable for whoever holds the licence, never part of a shared analysis.
In practice
Two views of the same rock. Both run in your browser on synthetic data — they show the machinery, not a deposit.
The rock divided into equal blocks, each with a grade. Raise the cut-off and the low-grade blocks disappear — what remains is the body worth mining.
Five holes from a surface, each split into intervals coloured by grade. A drill hole has a position, a direction and a depth — it is not a row in a spreadsheet.
Drag to rotate. Synthetic holes; the grades follow the same field as the block model.
Same blocks, two worlds
A block model and a neural network are the same thing seen from two directions: a set of values in space, and the connections you choose to draw between them. Our code moves between those two views constantly.
Drag the model to rotate it. The slider morphs the same set of blocks from a regular mine grid into a layered network — no blocks are created or destroyed on the way.
Illustrative model. No data leaves your browser — everything is computed locally.
Contact
Want to know more about the platform, discuss working together, or just see a demo? Do write — in Swedish or English.
Send an email