Devakrishna

I build software with AI agents and post what it costs and what breaks.

Now

One agent pipeline carries a topic from research to a written draft, through its own review pass, to a rendered video, with no hand-off in between.
I am cleaning up the skills that run it for open source: video production, layout QA, writing-style enforcement, and sandboxed web research.

Updated August 2026.

Ask the old code to review the new one. An AI rewrites a permissions function, nobody reads the rewrite, and the old code reviews it on a hundred real requests.
Run the shadow. Send the requests yourself, run both reads, break the shadow, replay the real round, and read the nine pairs that were the review.
The code I did not read. The AI rewrote a read, real requests ran both versions, and nine differing pairs were the whole review.
How I check a refactor I did not read. The old code keeps answering users, the new code runs beside it on real requests, and the mismatches per call site decide the switch.
A refactor I did not read. One read function, two versions on every live request, and nine differing pairs instead of the diff.
The issues with event sourcing after the prototype stage. Schema changes, projections, ordering, external replays, and plumbing, and what each one costs once the ledger is no longer small.
Same work, different judge, different winner. Two briefs written to tie, one LLM judge, and a verdict that tracked the judging setup, not the work.