AI Infrastructure Engineer. I orchestrate humans and agents with the same playbook — clear roles, zero ego, relentless accountability.
Previously Shopify · Atlassian · Bradfield CS
bharat@bharatagarwal.dev · Book a call · Gurugram, India
Knowledge graphs (5000+ nodes mapping curriculum dependencies), intelligent sales agents, multilingual video processing pipelines (24hrs→3min). Orchestration between business logic and frontier models.
Kubernetes autoscaling to 2.4M cores / 4.7M req/sec (Shopify BFCM). API redesigns cutting response times from 30s→9s. Chaos engineering and incident response at Atlassian.
Deterministic eval frameworks wired into CI/CD. Catching hallucinations, prompt regressions, and edge-case failures before they reach users. Built for non-deterministic outputs at production volume.
Every team member guards one zone — UX, metrics, evals, architecture. You can write anywhere, but you only guard one place. I borrowed this from Bill Belichick (six Super Bowls, one principle: “Do Your Job”) and applied it to agentic pods. It works the same whether the teammate is human or an AI agent.
Currently running Beads for agent memory and task coordination. Experimenting with Gas Town — multi-agent orchestration at the CLI frontier, 10–30 parallel agents sharing a task graph. The management challenge isn’t writing code anymore. It’s coordination at scale.
Building toward two patterns: ephemeral workers — session-scoped agents that spawn, execute, and dissolve — and deployed workers — persistent agents wired into infrastructure, running eval loops and monitoring pipelines without human intervention. The line between team member and infrastructure is getting blurry. That’s the point.
Built Mothership, an open-source PaaS. Chaos engineering and incident response at Atlassian, patent on a Confluence collaboration tool. Scaled Shopify to 2.4M cores for Black Friday. Computer science at Bradfield — networking, memory, concurrency, the layers beneath the abstraction. Then AI — knowledge graphs, agentic pipelines, and pedagogy-enriched EdTech systems.
Before all of that: five years of architectural training at SPA Delhi. Decomposing complexity into structure, articulating intent before touching material. In a prompt-driven world, that turns out to be an unfair advantage. The bottleneck is no longer building — it’s describing what to build.