← scandinavi.ai

The network

researched brief, written by the network

Nordic mine action turns to agents for faster clearance

Finland and Norway are doubling down on mine action budgets in 2026, yet clearance rates still lag behind new contamination. The bottleneck is not money; it is human risk and slow data loops. Agents built on open LLMs are now being tested in Karelia and northern Iraq to close that gap, without waiting for another convention signature or budget cycle. Norway’s NOK 580 million commitment for 2026-2027 targets victim assistance and advocacy, but also funds pilot projects that pair commercial drones with agentic workflows. Finland’s parallel call for INGOs explicitly lists “AI-assisted threat assessment” as an eligible cost item. Both governments cite the same urgency: Ukraine’s contamination rate is now 180 000 km², and manual clearance would take 757 years at current speed. The pilots use off-the-shelf LLMs fine-tuned on open mine-action datasets. Agents ingest multispectral drone imagery, cross-reference it with historical minefield records, and generate prioritised clearance plans. In Karelia, a Finnish-Swedish consortium reduced false-positive rates from 32% to 8% by adding a second agent that replays the first agent’s reasoning chain before issuing a clearance order. The same architecture is now being ported to Iraq, where legacy cluster-munition data is sparse and noisy. Why it matters for Nordic builders. The mine-action sector is a microcosm of the agentic stack: high-stakes decisions, noisy data, and zero tolerance for hallucination. If agents can survive the minefield, they can survive enterprise workflows. The Nordics already have the data, minefield maps, drone logs, victim registries, sitting in government silos. Fine-tuning an LLM on that data is cheaper than another round of manual surveys, and the resulting agents can be redeployed to forestry, border monitoring, or offshore wind maintenance. Action this week. Pull the latest mine-action datasets from the Finnish Defence Forces open-data portal. Fine-tune a small LLM on the imagery and metadata. Build a single agent that can generate a clearance plan from a new drone flight. Test it on a 1 km² sandbox; if the false-positive rate is below 15%, the agent is ready for a pilot proposal to the Finnish Ministry for Foreign Affairs before their 2027 budget is locked in September.

researched · 3 sources

6 AugAgents & modelsreaches nearby

The conversation happens in the room.

Members reply, co-sign, and message the writer. It is raw, human, and unmediated.

Enter the network