Subhash Kumar
thirty years in, still shipping · mumbai
I design enterprise AI platforms — the kind that have to survive production, not just a demo.
Thirty years from Visual Basic to LLM agents. These days I work where site reliability engineering meets agentic AI: systems that reason, plan, call tools, remember — and still answer to governance and an error budget. Lately that has meant designing agentic SRE systems for enterprise estates — autonomous triage, routing and recovery that have to convince architects and auditors, not just a demo audience.
System diagram
every architect draws themselves eventually · fig. 0
- 19,755 bookmarks
- chai, unmetered
- production incidents (unsolicited)
- model release notes, daily
Subhash v30.0
uptime: 30 years · unplanned restarts: 0
- agents that fix incidents
- this site — wiki, library, a curriculum
- opinions (unthrottled)
Now building
Agentic incident management
Autonomous agents that triage, route, and recover incidents — reasoning, tool calls, memory, and a policy engine, event-driven and observable end to end.
The coding-agent craft
Getting real work out of AI coding agents — and knowing which one to reach for. Claude Code as the daily driver: harness design, hooks, skills, subagents, and docs written for agents as much as humans. Benchmarked honestly against GitHub Copilot and Codex. And always the bill: token economics, cache behaviour, the right model for the job.
This site — an AI mentor in training
devopsmantra ingests my bookmarks, builds the wiki and the searchable library — and now teaches me back: a knowledge-graph curriculum, placement diagnostics, AI-graded recall and spaced repetition, two tracks live. 19,755 links in, the student became the syllabus.
The arc
Programmer analyst. Built software for whoever walked in the door — printers, perfume makers, financial firms, coaching classes, car-rental agencies. Visual Basic, FoxPro, SQL, and a lot of listening.
Enterprise consulting for European government and industry: document management, .NET, and a thirty-application portfolio that taught me estimation the hard way.
Team-lead years: audit reporting systems and insurance platforms. Where I learned that trust in a system is built from its reports — and that reliability was a discipline before it had a name.
Solutions architect for a very large UK retail estate. Architectural governance, migrations measured in millions, technology roadmaps that occasionally survived contact with reality — and a client hackathon won with an IoT build.
DevOps architect for a national-scale estate. Pipelines in Azure DevOps, infrastructure written down in Terraform instead of clicked into existence, monitoring wired into ServiceNow — the plumbing autonomous operations would later stand on.
The pivot. From keeping big systems boring — the highest compliment in SRE — to teaching them to fix themselves: LLMs and agentic AI, governed by policy, watched by telemetry.
Built an adaptive learning engine in public: knowledge graphs, AI-graded recall, spaced repetition. Learning agentic engineering the way I'd architect it — systematically.
Physics, University of Mumbai. Microsoft-certified across three decades — first in 2000, most recently on Azure. The tools change; the habit doesn't.
The obsession, measured
live from the vault · one dot = 25 saved links
Fresh in the vault
7 Jul jetrockets.com 7 Jul dotnet-presentations.github.io 6 Jul idratherbewriting.comSay hello
If you're building agentic systems that have to survive a real enterprise estate — or want to compare notes on getting serious work out of coding agents — I'd like to hear from you.
Not ready for hello? Start with the AI-agents wiki or wander the library.