vish engineering notebook
Screenprinted portrait of a hooded engineer in a visor

About vish Bengaluru-ish hours, mostly online

I am an engineer who ended up caring about the unglamorous half of AI: the runtime, the permission model, the audit trail, the evaluation harness, and the invoice at the end of the month.

Bio

I started out writing Android apps — MVVM, Room, Material, the whole modern-architecture checklist — and shipped a Jetpack Compose teaching project that a few thousand people used to learn the framework. Then I went through a Python-and-LLM phase, building small tools that made a model do something concrete: query JSON, execute JSON, chat with a PDF.

That phase is where the interesting problem showed up. The model was rarely the hard part. The hard part was everything around it: what the agent was allowed to touch, whether the result it reported was true, how much context it burned getting there, and whether any of it was reproducible tomorrow. That is the work I do now.

Today I work on agent infrastructure at ERP•AI. I own Proto, the desktop console where a business user directs agents, and build.host, the rail that lets an agent ship a site without a pipeline. Around both, I spend a lot of time on measurement: benchmarks, cache hit rates, router accuracy, cost per unit of work.

I am also stubborn about craft. This page is hand-written HTML and one stylesheet, because I wanted to remember what that feels like. Outside the terminal I shoot and cut short films, which is where I learned that attention is the only currency.

How I work

  • Understand the real flow first. Most proposed fixes are answers to a problem that does not exist yet. Read the callers before writing the abstraction.
  • Prefer the boring implementation. Delete confirmed-unnecessary code before adding new code. Reuse what is already installed.
  • Measure before and after. A refactor without a number is a preference.
  • Keep the denominator honest. Failed runs, timeouts and crashes count. A benchmark that hides them is a press release, not evidence.
  • Write it down while it is still annoying. The postmortem is the artifact; the fix is the side effect.

Track

2025 — Agent infrastructure at ERP•AI — Proto, build.host, evaluation, private model serving.
2025 — 2026 Product engineering across web and platform surfaces: CRM and enterprise app builds, public web properties, the publishing and hosting rails behind them.
Earlier Python and LLM tooling — JSON query and execution agents, document chat, small bots. Before that: native Android, Kotlin, Jetpack Compose.

Tools I actually use

Contact

The fastest way to reach me is email: vish@erp.ai. I am happy to talk about agent harnesses, evaluation design, self-hosted inference, or why your deploy tool lies to you.

Elsewhere: GitHub, X.