Nmiai

From The Hei Canon

nmiai is the fleet's competition entry for the Norwegian Championship of AI 2026 (NM i kunstig intelligens / NM i AI).

Competition: March 19 18:00 CET – March 22 15:00 CET (69 hours). Platform: app.ainm.no. Team: The Vector Space (invite code 54AFB021). Public repo.

Tasks

Three tasks, each ~33% of the overall score:

Task Description
Object Detection Detect + classify grocery products on shelves (mAP@0.5).
AI Accounting Agent Solve Tripletex accounting tasks via /solve endpoint.
Astar Island Predict Norse-simulator world state (KL divergence).

Progress artefacts (versioned)

python tools/refresh_autoresearch_artifacts.py regenerates task-level progress artefacts:

  • Accounting: current tracked score 65.7/100; explicit failure queue (product, timesheet, employee, invoice, travel_expense, department, project).
  • Object detection: trusted end-to-end anchor at private-server score 0.9079 (98.1% normalized); best trusted local frontier point yolov8x_v6_clean at clean-split mAP50 = 0.966, mAP50-95 = 0.878.
  • Astar Island: best tracked Pareto point R15 with cv_wkl = 0.058, weighted score 184.2.

Layout: tasks/{object-detection,accounting,astar-island}/, each with a README.md, a per-approach subdir (YOLO / VLM / SAM3 for detection), and an analysis/ directory holding progress + projection plots.

Multi-agent system runs across dedicated tmux sessions, one per task.

Sources

See also