Mikhail Parakhin

Mikhail Parakhin is a technology executive whose AI work spans search, digital assistants and commerce. He joined Shopify as chief technology officer in August 2024, overseeing its engineering and data organizations. In his April 2026 interview, he explained his preference for agents that critique one another, reproducible experimentation and customer simulations calibrated against historical behavior.[3][1]

Parakhin said Shopify effectively funded unlimited tokens for employees in April 2026, while encouraging strong models and allowing people to choose their tools. His qualification was direct: “One is that it’s not about just consuming tokens.” He favored a critique loop in which one agent produces work, another—ideally using a different model—reviews it, and the first revises. In his judgment, the additional latency bought better code than many isolated agents running in parallel.[1]

His concern was code volume: even if a strong model produced fewer bugs per unit of code than an average human, generating much more code could still send more bugs into production. He therefore emphasized the balance between spending on generation and spending on rigorous review. Shopify used its own review tooling, he said, because available products did not fit his preference for large models taking turns. Longer review could save overall deployment time by avoiding failed tests, removal of offending pull requests and rollbacks. He also credited work with Graphite on stacked pull requests, while describing repository interaction and CI/CD as major bottlenecks.[1]

Parakhin described Tangle as a collaborative system for data processing and ML experimentation. A team could clone a pipeline, change a component and carry the same workflow into production without porting it to another system. Content hashes allowed unchanged outputs to be reused even after a version change; shared preprocessing could run once for experiments across different teams. He placed Tangle after Ether and Yandex’s Nirvana in his experience, describing it as incorporating lessons from those systems and Airflow.[1]

Tangent added an agent-driven optimization loop: modify a pipeline, run experiments, measure a goal and repeat. Parakhin explained that it could rearrange existing components or inspect their code and create new ones, because Tangle components wrapped command-line calls and YAML files. He said this let product managers bring domain knowledge to experimentation without manually changing code. He credited Andrej Karpathy with popularizing autoresearch and Patrice Simard with sustained earlier advocacy for AutoML.[1]

Parakhin recalled testing autoresearch on a personal problem he believed had already been optimized over many years. After several weeks and more than 400 experiments, it found one improvement. He had expected the exercise to expose its limitations; the result made him a supporter. His enthusiasm retained a boundary: he found the approach effective at overlooked standard improvements, but less effective at wholly unfamiliar ideas requiring days of thought.[1]

Parakhin called SimGym initially his “brainchild” and recalled Tobi questioning whether simulated customers would simply repeat their prompts. His answer depended on Shopify’s historical records of storefront changes and sales outcomes: denoising and aggregation could supply signals for calibrating agents. He described an internal target of 0.7 correlation with add-to-cart events. The simulations used visual information and browser actions to represent the friction customers encounter; he credited Fireworks, CentML and NVIDIA with infrastructure optimization work.[1]

He also acknowledged a mistaken initial assumption. The team built a comparison between two storefront variants, thinking about merchants who lacked enough traffic for statistically significant experiments. As he put it, “What we didn’t realize is that most people don’t have A and B, they just have one thing”. SimGym subsequently evolved to examine an existing storefront and suggest changes intended to improve conversions. Parakhin said a store’s previous customer behavior helped tailor simulated agents to its particular customer distribution.[1]

Beyond storefront simulation, Parakhin described an HSTU-based system that modeled merchants as companies following possible trajectories. It could simulate paths with and without interventions, such as a coupon or campaign, and vary their timing. He identified this modeling of people, companies and possible interventions as a particular technical interest.[1]

Parakhin’s April 2026 account of Shopify’s use of Liquid AI emphasized workload requirements. He reported approximately 30 milliseconds end-to-end for query understanding with a roughly 300-million-parameter model, crediting optimization work with Liquid AI, CentML and NVIDIA. At the other end of the spectrum, he described distilling large models into task-specific Liquid models of roughly seven to eight billion parameters for catalog processing and Sidekick Pulse, where throughput and longer context mattered more than immediate response time. He praised the architecture while distinguishing these uses from frontier-model competition, and said Shopify continually tested alternatives.[1]

Before Shopify, Parakhin worked on search and assistants at Microsoft and Yandex. He headed Microsoft’s Bing Multimedia Search team from 2010 to 2014 and subsequently served as Yandex CTO, leading its technical teams globally. He left Yandex’s executive role in August 2019; by October 2020 he was Microsoft’s Corporate Vice President of Technology. He holds a master’s degree in physics from the Moscow Engineering Physics Institute. Shopify’s appointment announcement credited his Microsoft work on consumer and enterprise AI products including Copilot. Parakhin later described his previous Microsoft position as leading a business unit encompassing Windows, Edge, Bing and advertising.[4][3][1]

That earlier assistant work supplied context for his retrospective on Bing’s Sydney. Parakhin recalled deliberate personality shaping, drawing on lessons from Yandex’s Alice: Sydney’s character was partly designed rather than entirely emergent. He said the team sought politeness with a little edginess because it drew people in. He credited Andrew McNamara as Sydney’s first development lead and said McNamara subsequently led Sidekick and Pulse at Shopify.[1]