Field notes — the blog

Notes from inside the system.

Specific write-ups on decisions that moved the needle: architecture trade-offs, delivery pipelines, and leading engineers who are better than you at their part of it.

              {{ totalPosts }} notes
              {{ totalMinutes }} min total
              Monthly
            

/

                Latest
                {{ featured.date }}
                {{ featured.cat }}
                {{ featured.read }}
              

{{ featured.title }}

{{ featured.excerpt }}

Read the note →

// before → after
{{ row.text }}

{{ listHeading }}

{{ resultCount }}

                {{ post.cat }}
                {{ post.date }}
              

{{ post.title }}

{{ post.excerpt }}

{{ post.read }}

Nothing matches that.

Try a different term, or clear the filters.

The architect's dispatch

One note a month. Nothing else.

A single deep write-up on a system I worked on: the trade-off, the numbers, what I'd do differently.

            {{ article.cat }}
            {{ article.date }}
            {{ article.read }}
          

{{ article.title }}

{{ article.excerpt }}

{{ part.text }}{{ part.text }}{{ part.text }}

{{ b.alt }} — {{ b.caption }}

{{ b.label }}

{{ h.text }}

{{ cell.text }}

{{ b.label }}

{{ link.label }} {{ link.note }} {{ link.host }}

{{ b.caption }}

{{ b.text }}

{{ b.text }}

{{ b.label }}

{{ line.text }}

{{ item.no }}

{{ part.text }}

{{ part.text }}

{{ part.text }}

Takeaways

▖ {{ t }}

Tags

{{ tag }}

← Newer

{{ prevPost.title }}

Older →

{{ nextPost.title }}

Related notes

{{ post.cat }}

{{ post.title }}

{{ post.read }}

Shipping an Android app with an LLM in the loop

AI · Feb 2026 · 6 min read

Concept to Play Store with AI doing the scaffolding. What it genuinely accelerated, and the two places it quietly cost me a day.

I shipped a small Android app last quarter, mostly to find out where AI assistance actually helps on a codebase I did not already know by heart. The honest answer: it compressed the boring 70% and left the hard 30% exactly as hard.

Where it paid off immediately

Anything I could specify precisely came back correct on the first or second attempt. Room entities and DAOs from a schema description. RecyclerView adapters. Repository interfaces with the error cases enumerated. Unit tests for behaviour I could describe in a sentence. A working skeleton existed in an afternoon instead of a week.

Where it cost me a day, twice

First, a permissions flow that looked entirely plausible and did not match current platform behaviour on recent Android versions. It compiled, it ran, and it silently did the wrong thing on a real device. Second, a data layer that was correct until two writes overlapped, at which point it was quietly lossy.

Both of those were review failures, not model failures. I merged something I had read but not thought about.

The rule I settled on

I now treat generated code as a competent new hire's first draft: often good, sometimes confidently wrong, and never merged unread. The review bar is exactly the one I would apply to a person, with extra suspicion in two areas — anything platform-version-specific, and anything touching concurrency.

Specify tightly and it is fast. Specify loosely and it invents something plausible.
Platform quirks and concurrency still need a human who has been burned before.
The time saved is real, and it moves entirely into review. Budget for that.

Takeaways

AI accelerates the parts you can specify precisely; it invents plausibly for the parts you cannot.
Review generated code exactly as you would a new hire's first pull request.
Be most suspicious around platform-version behaviour and concurrent writes.

All notes · Shehzad Aslam