Competitions & Agent Ladders
How we test our thinking against public benchmarks, competitive multi-agent ladders, and hard computer vision problems.
Kaggriculture: Autonomous Farming Agent Ladder
A one-on-one multi-agent game simulating a full thirty-day farming season across a shared ten-by-ten board. Two autonomous agents compete under hidden information, resource contention, and fluctuating market prices.
Balancing early cash needs against late-game compound crop yields, while guarding your tiles against opponent starvation moves.
Spatial heuristic search backed by an offline self-play training arena and risk-adjusted discount horizons.
Greedy yield farming loses quickly to spatial containment. Building a strict local test harness was necessary to avoid ladder timeouts.
Scalable Handwriting OCR for Historical Surveys
Automating line segmentation and character recognition for degraded, multi-author handwritten agricultural survey logs.
Heavy variation in cursive handwriting styles, faint ink bleed, and missing boundary boxes across thousands of historical scans.
Hybrid line recognition models paired with morphological image cleanup, synthetic cursive generation, and domain vocabulary filters.
Careful image pre-processing (deskewing and adaptive contrast) improved the character error rate more than throwing bigger models at raw scans.
Competitive Battle Engine (Kraken AI)
A real-time competitive game AI built to play high-level matches under hidden movesets, predictive switching, and turn-order speed ties.
A sprawling decision tree caused by simultaneous turns, hidden items, and damage rolls with high variance.
Fast search with alpha-beta pruning, opponent intent modeling based on team-preview patterns, and dynamic damage thresholds.
Predicting switches beats raw damage math every time. Hard turn timers mean you must prune deep search trees aggressively.