All systems operational

Gaurav G S

Senior Site Reliability Engineer at Mastercard. Previously PhonePe, on UPI.

Four years keeping payment and banking systems up: on-call, migrations, disaster recovery, and the automation that makes those boring. Now I build open-source control planes that let AI take part in operations without taking the controls.

  • LLMs advise.
  • Deterministic systems decide.
  • Policy gates.
  • Audits prove.

Here's what that looks like on a live system

A small payments stack. Break it, then watch it heal itself.

Requests/s
1,240
p99 latency
182 ms
Error rate
0.02%
Error budget left
86.0%
  1. NoticeArgus
  2. ConnectArgus
  3. SuggestLLM
  4. ApproveJudikt
  5. FixHelios
  6. ConfirmAttesta

What happened hash-chained

  1. All quiet. Nothing has broken yet.

The incident is simulated in your browser. The way it gets handled is real: each step is one of my open-source projects, below.

Experience

Production reliability for systems where downtime means someone's payment didn't go through.

  1. Apr 2025 – now

    Site Reliability Engineer 2

    Mastercard, Pune

    Reliability for the Franchise and Legal Solutions platform (FLS/BRAM), used by global banks in a business line worth $77M+ a year.

    • Automated deployment and DBA validation workflows in Jenkins and XLR, so regressions surface before release and rollbacks are routine.
    • Lead incident triage by correlating Splunk logs, deployment state, database workflows and latency signals to cut time to root cause.
    • Led data migrations and recovered corrupted financial data during production incidents.
    • Run on-call, DR readiness, runbooks and post-incident reviews with development, DBA, security and compliance teams.
  2. Aug 2022 – Apr 2025

    Site Reliability Engineer 1

    PhonePe, Pune

    Bank PSP infrastructure for UPI, on a platform handling 360M+ transactions a day.

    • Migrated Yes Bank's PSP orchestration from Mesos to Drove, PhonePe's in-house container platform, with zero downtime.
    • Moved Axis Bank and Yes Bank infrastructure between data centers with zero downtime.
    • Parallelized MariaDB backup verification and ran it as code with SaltStack, making it 33% faster.
    • Built Nginx and ZooKeeper monitoring into Grafana for transaction-flow debugging and p95/p99 latency investigation.
    • Took 24×7 on-call, ran DR drills, and closed RBI compliance findings.
  3. Mar – Jul 2022

    Site Reliability Engineering intern

    PhonePe, Bengaluru

    • Worked on ZooKeeper and GlusterFS in production infrastructure before converting to full time.
  4. Dec 2020 – Mar 2021

    Cloud apprenticeship

    Google Cloud CR program

    • Completed the Architecting with Google Compute Engine specialization and 31 hands-on Google Cloud labs.

Blog

Write-ups from building these systems: what I tried, what broke, and what I'd do differently.

9 min read

I built Judikt: a security and reliability control plane for AI agents using MCP

What happens when an AI agent can restart production, query infrastructure, or execute operational tools?

Connecting an agent to a tool is the easy part. This post walks through how Judikt decides whether each call should run at all, using explicit policy, identity, risk scoring, approvals and rate limits, why tool responses are treated as untrusted input, and how a hash-chained audit trail proves what happened.

  • AI infrastructure
  • MCP
  • LLMOps
  • AIOps
  • SRE
Read on Medium

Projects

Four open-source systems that together heal the stack in the demo above. Judikt, the one that keeps AI agents in check, is published and documented.

  1. JudiktApproves

    Security gateway for AI agents calling tools over MCP

    Sits between AI agents and the tools they call. Enforces policy-as-code, OAuth/JWT identity, human approvals, rate limits and circuit breakers, and filters prompt and response injection. Every call lands in a tamper-evident, hash-chained audit log. Ships AttackBench for testing injection defenses.

    Python, MCP, OIDC/JWT, Redis, OpenTelemetry, Terraform, AWS

    pip install judikt v0.3.0 on PyPI Read the write-up on Medium
  2. ArgusNotices and connects

    Incident correlation and root-cause analysis, with AI kept advisory

    Ingests alerts, walks the service map to group them into one incident, and scores root causes deterministically before an LLM is allowed to explain anything. Fixes are typed, idempotent and policy-gated, and every decision lands in a PostgreSQL audit trail.

    Go, PostgreSQL, NATS, Redis, Prometheus, Grafana, Loki, OpenTelemetry

  3. HeliosFixes

    Durable workflow engine for running fixes safely

    Executes DAG workflows with PostgreSQL as the source of truth and NATS for dispatch. Worker leases, heartbeats, retries with backoff, timeout and dead-worker recovery, priority queues and backpressure, so a fix survives a crash halfway through.

    Go, PostgreSQL, NATS, Prometheus, Grafana, Kubernetes, Argo CD

  4. AttestaConfirms

    Kubernetes reliability control plane for safe rollouts

    Scores services for production readiness, checks SLOs, guards progressive rollouts and rolls back automatically when a release breaches them. Generates the Terraform, Argo CD, Kustomize, alerting and tracing config a new service needs.

    Go, Kubernetes, Argo CD, Terraform, Alertmanager, OpenTelemetry

Toolbox

What I reach for, grouped by the job it does.

Cloud and automation
AWS, GCP, Python, Go, Bash, Terraform, SaltStack, Ansible
Platform and delivery
Linux, Kubernetes, Docker, Jenkins, XLR, GitHub Actions, Argo CD, Nginx, HAProxy
Data
PostgreSQL, MariaDB, Redis, NATS, ZooKeeper
Observability and reliability
Prometheus, Grafana, Splunk, OpenTelemetry, Jaeger, Alertmanager, SLOs, incident response, DR drills

Education

2018 – 2022

B.E., Information Science and Engineering

CMR Institute of Technology, Bengaluru. GPA 8.4 out of 10.