Shopify¶
Commerce platform. Under CTO mikhail-parakhin, running one of the most aggressive internal AI rollouts in tech — daily-active AI tool usage approaching 100% of engineers by April 2026, with a sharp December 2025 inflection after model quality crossed a threshold.
Internal AI stack¶
- tangle — 3rd-generation reproducible ML/data workflow system; clone-an-experiment, one-click ship to production
- tangent — auto-research loop on top of Tangle; iterates code/prompts/infra toward a metric
- simgym — customer simulation environment (browser-based agents, buyer + merchant personas, driven by historical session data)
- ucp-catalog — product catalog as runtime infrastructure for agents (search, lookup, bulk API)
- liquid-ai — non-transformer models distilled from frontier for low-latency / long-context workloads (search, theme gisting, Sidekick Pulse)
Data moat¶
Years of real merchant + buyer session data is the moat around SimGym — see simulation-data-moat. Without historical traces, simulated customers only do what you prompt them to do; with them, agent behavior can be calibrated against reality and counterfactuals become meaningful.
Dev infra¶
Uses graphite for stacked-diff PRs. Parakhin flags git/PR/CI-CD as the highest-priority bottleneck in an agent-native world — code-generation volume has outrun the review/deploy pipeline, connecting to ai-generated-code-is-untrusted and agents-scale-not-fix.