Design a transformer-based TSP route optimizer

Customers send weighted complete graphs and want a near-optimal traveling-salesperson tour back. The team trains transformer models on generated graphs solved by a classical solver, and serves one model version online.

Design the whole loop: generating and versioning training and evaluation instances, training candidates, measuring route quality against a trusted baseline, promoting or rolling back a model reproducibly, and the inference path. Every answer must be a valid tour, so cover how you check the model's output and fall back to (or repair it with) a slower deterministic solver, and what the response reports about that.