Shopify

When I joined the team, I drove the strategic pivot to web components — and then helped form and execute the per-component rollout that shipped it across admin. The web component library itself predates me, but the migration strategy and its execution was the team’s work, with me leading.

• A shared coverage dashboard so every stakeholder — from team to CTO — looked at the same number.
• Per-team coordination (Slack, file links, preview URLs) before merging breaking migrations.
• Per-component campaigns owned end-to-end by Polaris engineers, with AI-assisted codemod tooling the team built and shared.
• A pragmatic incident posture when we caused regressions: owned communication, impact analysis, and the systemic fix.

A swarm-style platform that distributes large-scale frontend migration work across many coding agents in parallel. Each agent picks up a slice, generates a PR, validates it visually and semantically, fixes broken tests, and reports back. A coordinator stitches the output into a coherent campaign and hands off the long tail to humans where the swarm gets stuck.

Deployed on GCP Cloud Run with IAP/OIDC auth. Played a significant role in closing out the Polaris web component migration — the same shape generalises to any large-scale codebase change, and is where I think most enterprise frontend work is headed.

A suite of CLIs giving AI coding agents structured access to Shopify’s internal systems — CI, Slack, Vault, observability, BigQuery, the monorepo. Each tool is shaped to minimise round-trips and compose cleanly through bash, which agents already know how to use.

Adopted in 14 zones at peak, including Shopify’s in-house coding agent. In head-to-head testing against the equivalent MCP server, the CLI used 3× fewer tool calls, 60% fewer tokens, and finished in under half the time. That delta is why the project is shaped the way it is.

A personal memory system that watches what I do at Shopify — Slack I send and read, meetings I attend, docs I write and review, code I touch — and lets the agents I use reason against that memory the way a good colleague would. Not “what did I literally say on March 4th” but “what’s been decided about X,” “what does Pulkit usually push back on,” “what did we conclude about the egress fix.”

Built on Hindsight (open-source agent-memory engine), runs entirely on Shopify cloud infrastructure, single-tenant by design. The interesting work is in the four layers above raw vector search: structured facts extracted at retain time (not raw text); tag conventions as a soft schema (visibility, project, decision_type, person); a single bank with visibility as an egress filter rather than three siloed banks; and observations — a background worker that consolidates evidence across many memories into deduplicated beliefs with proof counts. That last layer turns Thoth from “I can find that conversation” into “I know what Pulkit will say.”