Durgapur, India

Deb.

Cloud, DevOps & SRE in the making

I learn by owning things that break. Everything I run sits on a private network with no port open to the internet, behind a single identity provider, with backups I have actually restored from.

Deb, the author of this site
Me, presumably. The only photo here is one I actually took.
Studying
BCA
Learning
Go, Python
Reading
SRE blogs at 1am
Looking for
SRE / DevOps internships

About

The short version

I prefer understanding the low level before trusting the high level.

The stack behind this site is deliberately boring. The constraint of having no public address to hide behind is what taught me the most — every service has to justify its own exposure, and none of them have any.

I'm early in this, and I'd rather be honest about that than pad a résumé. What I can do is take a system end to end, work out why it broke, and write down what I changed so the next failure is cheaper than the last one.

  • Linux & shell

    Daily-driver Fedora and Debian hosts, encrypted at rest. I am comfortable breaking my own systems and reading the logs that explain why.

  • Networks

    Mesh overlays, residential NAT, dual-stack path selection, split-horizon DNS. I test the path instead of trusting that it works.

  • Containers

    Compose stacks where each service carries its own tunnel sidecar, so publishing something is a deliberate act rather than an accident.

  • Identity & access

    One OIDC provider in front of everything, passkeys over passwords, closed registration, and secrets kept out of version control.

  • Automation

    Backups, health checks and deploys written as scripts I can read in one sitting. If I cannot reproduce it, I do not trust it.

  • Reliability

    Restore drills, MTU discovery, packet-loss budgets, and logs that record what happened instead of what should have.

The work

Problems, not products

Nobody hires a list of software. These are the problems I ran into owning infrastructure, and what I did about them.

  • Everything I run was one port scan away from being somebody else's problem.

    A network with no front door

    Every service I run sits inside a WireGuard mesh and is reachable only from my own devices. The machine hosting them admits two ports — SSH and the tunnel itself — so there is nothing for a scanner to knock on. TLS terminates inside the mesh, which means no certificate renewal on a box that faces the internet.

    • WireGuard mesh
    • Tailnet lock
    • Mesh-issued TLS
    • Peer relay
    • Exit node
  • Password reuse across a homelab is still password reuse.

    One identity, one place to revoke

    A self-hosted OIDC provider fronts everything I run, so leaking one service's password does not leak the rest of the stack. Passkeys instead of passwords, signups closed, and a single control point if I ever need to cut access.

    • OIDC / OAuth2
    • PKCE
    • Passkeys
    • Closed registration
  • A backup I had never restored was just a rumour.

    Backups I actually verify

    Nightly snapshots to object storage, driven by a timer I can read in one sitting. Mutable state is deliberately excluded and databases are dumped instead of copied. The part that matters is the restore — I run them, because an untested backup is a hypothesis.

    • restic
    • S3-compatible object storage
    • Database dumps
    • Restore drills
  • My ISP hands every customer the same address, and half my traffic stalls.

    Making a CGNAT line usable

    Residential NAT meant inbound connections were simply not an option, and outbound IPv4 fought with egress that lived in another country. I moved the workload onto a dual-stack mesh, proved the IPv6 path with throughput and MTU tests instead of assuming, and put the exit node wherever the NAT wasn't.

    • CGNAT
    • IPv6
    • Path MTU discovery
    • Exit node routing
    • Happy Eyeballs
  • Search engines keep a log of you. Mine keeps nothing.

    Search I can read the source of

    Self-hosted metasearch with a cache layer in front. No query log leaves the machine, no third-party script loads into the page, and I can read exactly which upstream engines are answering and which are quietly rate-limiting me.

    • SearXNG
    • Valkey
    • Docker Compose
  • A static site should not need a server.

    This site has no origin

    Built with Astro and compiled down to static files, served from the edge as a Worker. No runtime, no origin to scale, and almost no JavaScript — the whole page ships a few kilobytes of script, because the animations are CSS.

    • Astro
    • Svelte islands
    • Tailwind CSS
    • Cloudflare Workers
    View

Toolkit

What I reach for

  • Operating systems

    • Fedora (daily driver)
    • Debian
    • systemd
    • Btrfs + LUKS
  • Shell & scripting

    • Bash
    • Zsh
    • Git
    • systemd timers
  • Containers

    • Docker
    • Docker Compose
    • Sidecar networking
    • GPU passthrough
  • Networks

    • WireGuard mesh
    • CGNAT & dual-stack
    • Path MTU discovery
    • Split-horizon DNS
  • Identity & security

    • OIDC / OAuth2
    • Passkeys
    • Least-privilege firewall rules
    • Secret management
    • Mesh device approval
  • DevOps & reliability

    • CI/CD
    • Docker-in-Docker runners
    • Health checks
    • restic backups
    • Restore drills
    • Structured logging
  • Databases

    • PostgreSQL
    • Redis / Valkey
    • SQLite
  • Languages

    • Bash
    • C (coursework)
    • Java (coursework)
    • Python (learning)

Contact

Get in touch

Open to internships and junior SRE, DevOps or cloud roles. The fastest way to reach me is email; I'm happy to talk about anything I've broken.