NM i AI 2026 — Main Round

Solution code for the main round of NM i AI (the Norwegian AI Championship) 2026. Three independent tasks were tackled solo: a NorgesGruppen VLM task, a Tripletex accounting-API automation task, and the Astar Island probabilistic map-reconstruction game.

The qualification round (a separate MAPF warehouse-coordination problem) is not in this repo — this is the main-round work only.

Tasks

tracks/ml/ — Astar Island (probabilistic map reconstruction)

You get a limited query budget to peek at viewport observations of a hidden grid, then must submit a full probability distribution over 6 terrain classes for every cell. Submissions are scored by KL divergence against the true map: score = 100 · exp(−KL_weighted), with KL counted only over dynamic (non-static) cells.

Approach — a Bayesian observer (predictor.py):

• Dirichlet smoothing over per-cell class counts; α = 2.0 chosen by an offline score sweep (α=0.5→91.3, α=1.0→94.2, α=2.0→96.7, α=3.0→97.8 marginal).
• A probability floor (PROB_FLOOR = 0.001) on every class — a single 0.0 produces infinite KL and tanks the score.
• A two-phase observation policy (observation.py): Phase A spends budget on systematic coverage, Phase B targets the highest-uncertainty cells.
• A central orchestrator (main.py) running the budget→observe→predict→submit loop with continuous polling and per-seed submission.

Iteration is visible in the versioned participate_v*.py files; simulator.py is a local scorer used to tune α and the observation policy offline before spending real query budget.

tracks/cv/ — NorgesGruppen VLM

Vision task on NorgesGruppen retail imagery: a YOLO (Ultralytics) detector paired with a DINOv2/ViT embedding + reference-matching pipeline for product/shelf recognition. Training helpers (scripts/) target a single L4 GPU on GCP.

tracks/llm/ — Tripletex automation

LLM-routed handler for the Tripletex accounting task: classify an input, then dispatch it to the matching structured-extraction route (employees, invoices, projects, expenses, accounting entries).

Layout

• tracks/ml/ — Astar Island Bayesian predictor
• tracks/cv/ — NorgesGruppen VLM detection + embedding pipeline
• tracks/llm/ — Tripletex routing / extraction
• scripts/ — GCP training helpers (set GCP_PROJECT; no project IDs baked in)
• submissions/ — generated artifacts (gitignored)

Running the Astar Island predictor

pip install -r requirements.txt
export ASTAR_TOKEN=...        # competition JWT — read from env, never hardcoded
python tracks/ml/main.py --token "$ASTAR_TOKEN" --poll

Notes on this repo

• No credentials are committed. Competition token is read from ASTAR_TOKEN; GCP scripts read GCP_PROJECT from the environment.
• Live working notes, API dumps, and internal endpoints from the competition were removed before this repo was made public.

License

MIT — see LICENSE.