· Updated October 25, 2026 · 3 min read

Starting over, on purpose

A fresh open notebook and pencil inside a dashed boundary with a check mark, beside a stack of older closed notebooks
On this page

The questions I’m working on have changed, but this site still reflected my earlier work. Writing is how I think, so I wanted a place to think the new questions through.

This site is that reset. The previous version came from a different chapter: deep dives into Scala, Spark and distributed data systems. That work shaped me. But the questions I’m working on now are different.

What changed

For years I worked at the infrastructure layer: pipelines, lakehouse architectures, distributed compute, performance tuning.

The goal was always the same: move data from one end to the other and transform it exactly as written. A schema mismatch, a resource limit or a skewed partition could break a pipeline. The intended transformation was explicit, even when debugging was hard.

Then I started asking a different question. What if a pipeline could interpret its data? Notice an anomaly, decide where something should go, and recover from an ambiguous case without someone writing the recovery logic first.

For me, it felt like moving from writing a checklist to handing a task to a coworker. I had to decide what they could decide on their own, and when they should ask for help. That question drew me towards agents, where a model proposes the next step instead of every step being written in advance. It raises questions I didn’t have to ask before:

  • How do you give a system autonomy without losing control of it?
  • Where should a person stay in the loop?
  • What does reliability mean when the answers can vary?

Those used to sound like research questions. Now they’re product questions.

What I’ll write about

Three areas keep pulling me back.

Agents in production

Systems that survive ambiguity, failures and real users, not demos. I want to write about how agents are put together, how you evaluate them, and what it takes to trust one with real work.

How people and agents work together

I want to explore interfaces that help people decide what to do next: approve an action, correct a mistake or ask for more information. My question is whether that kind of guidance means people spend less time supervising an agent.

Building with judgement

The longer I build, the more I care about the questions before the code. Is this the right problem? What’s the smallest version worth shipping? Should we build it or buy it? When is it good enough?

AI can make a first version much faster to build. My worry is that it also makes a weak idea look convincing before anyone has tested whether people need it. It makes it cheaper to build the wrong thing convincingly. A mediocre feature built fast with AI can look polished enough to pass review. When building gets easier, deciding what not to build becomes the real constraint.

Why rebuild the site

Rebuilding is easier than ever, but speed isn’t why I did it. I wanted the publishing workflow to match how I think: idea, hypothesis, pressure test, refine, publish. Fewer posts, each one an argument I can defend. If I can’t defend it, it doesn’t ship.

Why “gcdaii”

The site is a reset, but the name comes from an old nickname: my initials plus a small personal detail. In Nepali, dai means “older brother”, and friends have called me “gcdai” for years. I added the extra “i” because the handle was free.

If you work where data, AI and product meet, I’d like to hear from you.

Learning never stops.

continue reading

Your code survives the session. Your reasoning doesn't.

A practical guide to a personal journaling system for AI-assisted development: a Claude Code skill and a hook that capture the reasoning your code can't.

Building the AI-ready data platform

Capable agents can still fail on stale or ambiguous data. What an AI-ready analytical platform needs, and why I favour a lakehouse foundation.

Data is the bottleneck in the agentic era

Agents need dependable data, and they create records of their own. Why I think data delivery becomes the constraint once the model can do the job.