- krishna-cp
- arjuna-1
- arjuna-2
- monitoring
- grafana
- prometheus
- loki
- alloy
- hermes
- immich
- calorie-vision
- splitwise
- vaultwarden
- seafile
- jellyfin
- navidrome
- audiobookshelf
- silverbullet
- karakeep
- tt-rss
- postgres
- valkey
- proxy manager
- tailscale
Hover a thread: each one is a service running at home.
Vishal Dodda
SRE for blockchain infrastructure. I work where reliability, AI and web3 cross. Follow one thread at a time; by the end of the page you’ll see where they meet.
I never treat a technology on its own.
The wonders happen where two of them meet, so that's where I build.
Woven at home
My data lives on machines I own.
My photos, passwords, files, notes and media run on two Proxmox machines at home: around 25 self-hosted services, including the tools I write for myself and my own AI assistant. A four-node Kubernetes cluster and a Tailscale mesh tie them together.
Privacy is the point. Nothing reaches them except through a Cloudflare Tunnel, so not one port is open at home, and the most personal parts, like my assistant's memory, never leave the network. It's also where I break things before anyone pays me to fix them.
Running it every day taught me my trade: keeping things up.
- immichselfhosted-trixie
- hermes agenthermes
- silverbulletselfhosted-trixie
- prometheusmonitoring
- calorie-visionbuilt by me
- kuberneteskrishna-cp
Reliability
Keeping systems up is my craft.
Metrics, logs and the patient work of knowing why something broke. I practise on a four-node Kubernetes cluster I built by hand with kubeadm, where every node pushes its metrics and logs to one monitoring node.
When the dashboards aren't enough, I go down to the packets themselves, in the kernel.
Today I practise that craft on blockchains.
0 screened0 dropped in kernel0 sent up to userspace
XDP attack filter
An eBPF program at the network interface that screens HTTP for XSS, SQL injection and traversal, backed by an ML ensemble. It sends 403× fewer packets up to userspace.
Kubernetes, by hand
A four-node kubeadm cluster with Flannel, spread across two Proxmox hosts over Tailscale: krishna-cp drives arjuna-1, arjuna-2 and a monitoring node. Grafana runs inside it.
Simple Search
A 3-tier search app on the Perplexity API, with frontend, backend and database each containerised and deployed on Kubernetes, logging every search.
Homelab observability
Alloy agents push to Prometheus and Loki on a dedicated monitoring node; Grafana runs inside the cluster. Push-based, so there is no up{} series to lean on: health checks query what each agent actually sends.
Chains
By day, that craft runs blockchains.
I'm an SRE at a stealth, fast-paced web3 validator startup, working on Solana and EVM chains and the RPC infrastructure behind them.
A Jacquard loom reads its punch cards strictly in order, each laced to the next; a chain works the same way.
What I add to that work is AI, starting with the agents I live with.
RPC and chain operations
Day-to-day SRE work for Solana and EVM chains, including the RPC infrastructure applications talk to.
Westend validator
A Substrate validator on the Westend testnet, with proxy-account staking, session-key rotation, and chill and unchill drills, run as a dedicated service user.
ML credit-scored lending
A DeFi lending and insurance dApp where an off-chain ML credit score is attested on-chain with EIP-712 signatures and sets the borrower's risk tier.
Agents
Day to day, I live with the newest of AI, inside a privacy boundary I drew myself.
My assistant runs on Hermes, an open-source agent that writes and refines its own skills, runs cron jobs and spawns subagents for parallel work. Its long-term memory is Honcho, which builds a working model of me. Both run on one VM in my homelab, so the reasoning never leaves my network.
My second brain lives in SilverBullet, which Hermes reads and writes over MCP; Obsidian stays isolated, mine alone. Local models come from Ollama, and when a job needs a bigger one, it goes out through OpenCode Go's zero-data-retention endpoint. Alongside it, Paperclip runs a small company of agents with an org chart, goals and budgets.
Beyond the agents I use, I build models of my own.
My assistant stack as I drew it. Pick a node to see what it does and where it runs.
Private AI assistant
Weeks of building an assistant around privacy: a self-hosted Hermes agent with Honcho memory, a local LLM, personal skills and cron jobs that explain things the way I learn them.
Paperclip agent company
Paperclip, self-hosted: agents arranged in an org chart, each with goals and a budget, and every action in an audit log.
Daily DevOps agent
My first agent: every morning it gathers news from RSS feeds, picks ten DevOps questions and a quote, and emails me a digest, scheduled by cron.
Models
And I build models that do a real job.
The one I'm proudest of lets you ask a Postgres database a question in plain English. An agent plans and validates the SQL over MCP, but it never executes a write and never sees the rows: results flow straight from the database to your screen.
My favourite to watch is logSonar. Dashboards never gave me a feel for a node, so I turned logs into pixels: each log line is weighed down to three values from 0 to 255, one RGB pixel, and a batch of them becomes an image per service. KMeans names the mood, FFT finds the slowdown that keeps coming back, and anything 20% off the mean is flagged.
I'm studying Computer Science with a specialisation in AI and ML, and I like models with a job like that: classifying a signal, filtering a message, answering a question about a database.
As a research assistant in 2025 I balanced cloud, fog and edge, using multi-objective optimisation to move traffic off the cloud without costing the edge its latency.
Here's how those threads were laid, one row at a time.
Raw metrics pour in from three services and land as pixels, the way logSonar draws them. Synthetic logs, real method. Pick a quadrant to see another model.
Agentic Postgres over MCP
Ask a database a question in plain English. The agent plans and validates SQL through MCP, but never executes a write and never sees the result rows; data flows straight from Postgres to the browser.
logSonar
Logs translated into pixels: each log line becomes an RGB value, so a service's health reads as an image. KMeans names the mood, FFT finds the repeating slowdown.
Radar signal classifier
An ML classifier separating radar signal from noise, built during a defence research internship and deployed into the tool its engineers already used.
Hinglish spam detection
SMS spam filtering for code-mixed Hindi-English text, with its own transliteration and stopword pipeline. The SVM reaches 98.8% accuracy and 0.965 F1 on 606 held-out messages.
calorie-vision
Snap a meal and a vision model estimates calories and macros per item; scan a barcode for a Nutri-Score. It runs on my homelab.
The cloth so far
Each row is a stretch of work, newest on top.
Site reliability engineer
A stealth, fast-paced web3 validator startup
Solana and EVM chains, RPC infrastructure.
Research assistant, AI
VIT Vellore
July to December. Multi-objective optimisation that spreads outgoing traffic across fog nodes, easing cloud load while keeping edge latency low.
Intern
DRDO
Machine learning for radar signal classification.
Core committee and student coordinator
IEEE-CS · graVITas & Riviera
Hackathons and cryptic hunts, and a fest team of 60 serving 1,000+ attendees.
B.Tech, Computer Science (AI & ML)
VIT Vellore
Where the threads started.
Now watch all three threads cross.
- reliability
- models
- chains
- xdp filter
- logsonar
- observability
- rpc ops
- validator
- mcp agent
- hermes agent
- radar classifier
- hinglish svm
- calorie-vision
- credit dapp
- argmaxvishal
SRE × AI × Web3
The philosophy from the top, made literal: every project you just read is a node in this network. Reliability, models and chains go in, and its argmax, the single most likely answer, is the engineer I am: one who keeps chains running and puts models to work on real infrastructure.
Pull a thread
That’s the whole cloth: an SRE who keeps chains running, brings models into the work, and builds it all at home first.
If any thread caught your eye, pull it. I’m quickest to reply on LinkedIn, and I write now and then on Hashnode.