Shopify CTO Mikhail Parakhin appeared on the Latent Space podcast, published April 22, 2026, to discuss what AI adoption looks like inside Shopify. The conversation covers the company’s internal AI adoption curve, three internal systems — Tangle, Tangent, and SimGym — Shopify’s use of Liquid AI’s non-transformer architecture, and Parakhin’s view that review and CI/CD have become the binding constraint on AI-assisted engineering, not code generation.
Parakhin came to Shopify from Microsoft, where according to the Latent Space transcript he held an executive role.
The adoption curve and the review bottleneck
The Latent Space transcript describes December 2025 as an inflection point in Shopify’s internal AI adoption, with usage reaching what the transcript characterizes as a runaway trajectory driven by model quality improvements. Nearly 100% of daily active employees now use at least one AI tool, according to the transcript. PR merge growth reached 30% month-on-month, up from 10% previously, the transcript states. Shopify provides what the transcript describes as “unlimited tokens” with a minimum model threshold of Claude Opus 4.6.
Parakhin’s stated position, per the transcript, is that the real bottleneck is no longer code generation but review, CI/CD, and deployment stability. He argues that AI coding can still lead to more bugs in production even if models write fewer bugs on average than humans: the mechanism is volume — more code arriving faster creates more surface area to review, and if the review process does not scale, regressions reach production. The transcript indicates Shopify built its own PR review flow as a result.
The transcript also attributes to Parakhin the view that Jensen Huang is “directionally right” on token budgets but that raw token count is the wrong metric for evaluating engineering output.
Tangle: reproducible ML workflows
Tangle is Shopify’s internal system for making ML and data workflows reproducible and collaborative. The key design element the transcript highlights is content-addressed caching: outputs are keyed to inputs deterministically, so repeated work across teams is automatically deduplicated. The transcript notes a comparison to Airflow is made, with Parakhin arguing Tangle is different, though the network effects from content-addressing are described as central to its value.
Tangent: automated optimization loops
Tangent is Shopify’s auto-research system, used to optimize search, themes, prompt compression, and storage through automated experimentation loops. According to the transcript, Parakhin describes AutoML as “finally feeling real in the LLM era.” The transcript also notes that Tangent improved search throughput from 800 QPS to 4,200 QPS through optimization loops, and identified redundant datasets worth what the transcript describes as billions in bytes of storage. The transcript describes Tangent as accessible to product managers and domain experts, not only ML engineers.
SimGym: customer simulation
SimGym is Shopify’s customer simulation system. The transcript presents Parakhin’s argument that simulated customers only work if grounded in real historical behavior, and that Shopify’s transaction data across millions of merchants and buyers makes the system difficult for others to replicate. The transcript notes SimGym achieved a 0.7 correlation target with add-to-cart events in real A/B tests.
The system, according to the transcript, evolved from comparing A/B test variants to prescribing changes to a single live storefront to raise conversions, and to modeling counterfactuals about interventions such as discounts, campaigns, and notifications. Infrastructure costs are described as significant, with the transcript mentioning multimodal models, browser farms, and serving and distillation costs. The transcript also notes that Chinese Restaurant Processes appear in practice for clustering customer behavior.
Liquid AI in production
Shopify is using Liquid AI’s non-transformer architecture in production. The transcript describes Parakhin calling it the first genuinely competitive non-transformer architecture he has encountered in practice. Use cases cited are low-latency query understanding running at sub-30-millisecond latency, large-scale catalog work, and Sidekick Pulse workloads. The transcript describes Shopify’s model selection approach as not committed to any single architecture, and raises as an open question whether Liquid AI could scale to frontier level with sufficient compute investment.