Hi, I'm Sreekesh.

The closer I've gotten to the work, the more I've become a builder first and a manager second. I run the whole loop myself: gathering requirements and running delivery as an engineering manager, designing the system and its cloud architecture as an architect, and writing the production code as a full-stack developer. Owning the problem and shipping the solution is what I optimize for.

Around 2023 I started treating AI the way I treat any tool: hands-on, pragmatic, and measured by whether it ships working code. That became a conviction. The most effective teams will stop treating AI as a code generator and start wiring it into the loop, with intent capture, closed feedback and verification built into the workflow itself. Today my focus is agentic AI systems for software engineering.

• A sandboxed AI agent for report scripting at Cambium Networks. It turns natural-language requests into reusable scripts without ever touching production credentials.
• cnMaestro, monolith to microservices. Independently deployable, event-driven services on Node.js, SQS and Lambda, decomposed highest-churn modules first so releases stopped needing full-platform regression cycles.
• Real-time dashboards, about 85% faster. WebSocket channels instead of polling, with visible widgets rendered first: load times went from 30–40 seconds to under 5.
• Data at scale at QBurst. A Node.js pipeline processing 1TB+ a day, and an Elasticsearch layer that replaced a MySQL full-text bottleneck with sub-second search across millions of records.
• Teams that ship. Led up to 15 engineers through platform-scale shifts, and took sprint completion from an inconsistent baseline to a sustained 95%.

Method

End to end: requirements, architecture and code are one job, so nothing gets lost in handoff.

Defaults

Intention and closed loops. Whatever gets automated must also be verifiable.

Language

Direct, no fluff.

Measured by

Working code and outcomes: sprint completion, latency, cost, availability.

• Browser tools for terminal agents. Giving OpenCode visual awareness and MCP tools to test, inspect and debug web apps on its own. Read the article →
• Sandboxed agent infrastructure. From one container to an air-gapped, gVisor-sandboxed, cluster-scale agent architecture. Read the build →
• Closed-loop AI SDLC. Turning linear pipelines into loops: intent anchors, feature journals and validation against live specs. Read how →
• Agent-first products. Nirmanam, a construction ledger that an LLM agent can stand up for a non-technical owner from a short conversation. See the project →

• Intent over output. AI-generated code needs a source of truth. A simple feature.md keeps agents honest about what they're building.
• Close the loop. Whatever you automate, you must be able to verify. Feedback loops, not pipelines.
• Ship on time. Own the work end to end so nothing gets lost between people.
• Measure by outcomes. The numbers tell you whether the structure is right.

Outside work I bring the same builder's mindset to things that have nothing to do with software: family, curiosity about cultures, and the open-ended tinkering that keeps a 14-year career feeling like a hobby. I write about some of it in notes, and in English and Malayalam on my personal blogs.