About

I'm a backend engineer in Delhi. I build the platforms AI coding agents run on, and the controls that let a company trust them with real code.

I'm a software engineer at Fyscal, with a B.Tech in computer science from BPIT. My days go into OneXO, a multi-tenant platform where teams run coding agents in isolated sandboxes, inside budgets, guardrails and audit logs they control.

The other half is Go services for a financial platform. Routing, retries, fallbacks, and the traces that tell you which one fired.

  • IsolationAn agent should be able to run anything and break nothing. Every session gets its own microVM.
  • ReviewabilityWork ends as a pull request a person can read, never as a change that just appeared.
  • Boring reliabilityTimeouts, retries and fallbacks first. Features after.
  • EvidenceTraces and numbers before opinions. If it isn't measured, it isn't faster.
  1. 1RequirementA ticket or a spec, written by a person.
  2. 2PlanSmall spec modules with a decision log · plan-out
  3. 3SandboxOne microVM per session, up in about 0.3 s · claude-code-microVM
  4. 4AgentClaude Code, Codex or OpenCode writes the change and runs the tests.
  5. 5Pull requestEvery request ends as something reviewable, not a hand-off.
  6. 6ReviewA person approves it or sends it back.
AI
  • Claude Code
  • Codex
  • OpenCode
  • Mastra
  • AI SDK
  • MCP
  • Bedrock
Languages
  • Go
  • TypeScript
  • Python
Backend
  • Bun
  • Node.js
  • gRPC
  • Kong
  • Django
  • Redis
  • PostgreSQL
  • ClickHouse
Infra
  • AWS
  • KVM microVMs
  • Docker
  • Grafana
  • OpenTelemetry
  • Jaeger
  • BPIT, GGSIPUB.Tech, Computer Science and Engineering2022 – 2026

I send small fixes to the tools I use: Supabase, freeCodeCamp and Airbnb's style guide so far. See the pull requests

Email is best. I read everything and reply to most things.

Delhi · IST

Copied prateekbisht04@gmail.com