Agero

Roadside assistance isn't just a service call—it's a massive retention engine. See the latest data from Bain & Company on how a seamless rescue drives a 78% likelihood to renew.

From days to hours: how Go transformed my ML training pipeline I'm building a card game AI for Pepper, a 6-handed bid euchre variant played with a double euchre (pinochle) deck. The training loop is simple: simulate millions of hands, record every decision point, train an MLP to predict outcomes, repeat. The problem? When everything was in Python, a single training iteration took days. Not the training itself (PyTorch handles that fine), but generating the data. Each hand requires 20 counterfactual rollouts per decision point, across 6 players, 8 tricks per hand. For 500K hands with 20 rollouts per decision, that's roughly 40 billion individual card play evaluations per training iteration. So I rewrote the simulation and data collection in Go. The results were dramatic. What made the difference: Parallelism was the easy win. 8 goroutine workers collecting hands concurrently with minimal coordination. But the real gains came from profiling and eliminating allocations in the hot path (I used Claude Code to help identify bottlenecks and iterate on the optimizations): - sync.Pool for hand buffers and trick objects. Each rollout was creating and discarding dozens of slices per hand. Pooling dropped GC pressure by 90%. - In-place card removal instead of allocating new slices. One function change, measurable impact across billions of calls. - Flattened weight matrices (row-major []float32 instead of [][]float32) for cache-friendly MLP inference during rollouts. - cgo integration with Apple's Accelerate framework (CBLAS) for the neural network forward pass. The MLP runs inside every rollout, so even small per-call savings compound massively. Added a pure-Go fallback with build tags for Linux/Docker deployment. The numbers: Python collection for 300K hands: ~2-3 days Go collection for 500K hands: ~4-5 hours That's roughly a 20x speedup for 1.7x more data. A full self-play iteration (collect data, train card model, collect bid data, train bid model, evaluate) now completes in about 6 hours. I can run multiple iterations per day instead of one per week. Training stays in Python/PyTorch. It's the right tool for gradient descent. But data generation is pure compute with no framework dependencies, and that's where Go shines. The AI currently scores +1.73 pts/hand against a strong rule-based opponent after 8 iterations of self-play. Not bad for a 128x64 MLP. I'm curious if others have made a similar shift. Has anyone else moved their simulation or data generation layer out of Python into Go, Rust, C++, or something else? What was the bottleneck that pushed you, and what kind of speedups did you see? Code: https://lnkd.in/gcwafMte