I am a Senior Data & AI Engineer based in Kathmandu, Nepal. The job arrived in that order. First Python services and PostgreSQL schemas; then the databases those services leaned on, which is where I learned that most outages are a storage decision made two years earlier; then the platform around them — Kafka ingestion, Spark processing, Airflow orchestration, lakehouse modeling, analytical serving.
Agents came last and belong to the same stack. The interesting part was never the orchestration framework. It was the boundary: what a model is allowed to write, what validates it, what happens on the second attempt, and how you compare today’s answer to last week’s. I work close to system internals — throughput and failure limits, recovery paths, observability, and cost before scale turns an architectural shortcut into an operational one.
Operating principles
Ownership before consumption
Every published dataset needs a freshness expectation and a named owner before it gets a consumer. An unowned table rarely fails loudly; it keeps serving yesterday's numbers until somebody makes a decision with them.
Choose technology with an exit
I look at operating cost, data portability, and recovery work before I look at a feature matrix. A tool earns its place when the team understands both how to run it and how to leave it.
Agents produce records, not theatre
A chat transcript cannot be joined, audited, backfilled, or compared with last week's decision. If a model produces something worth acting on, it crosses a schema and validation boundary like any other producer.
Narrow contracts beat wide promises
Much of data-quality work is downstream cleanup caused by an upstream table promising more than it can hold. I would rather publish three columns with guarantees I can defend than a wide table everyone quietly stops trusting.
What I am doing now
At UXCam I work across the data platform and the stores underneath it: processing, serving, database reliability, governance, and architecture review. I also own the production agent workflows — LangGraph and Google ADK for orchestration, MCP for the tool surface — with the same requirement as any other producer: what they emit lands in the governed stores, typed and reviewable, or it does not land.
Outside that I keep a multi-engine failover lab with restartable scenarios, published commands and measured outcomes, and I build an independent product where I own every layer from the schema and API to the deployment.
Updated August 2026.
Background
I studied computer science at JNTUA College of Engineering in Anantapur, India, from 2015 to 2019. My first roles centered on backend services; the data half of the job gradually became the job. I speak English at C1 and Nepali natively.
I work from Kathmandu and shift my day later to overlap with European afternoons and US mornings, a rhythm I have used with distributed teams for years.
For occasional focused advisory or implementation work, consulting scope and availability are here.