Agentic AI architect and forward deployed engineer
Most AI pilots never reach production. I build the ones that do.
I work forward deployed: inside your company, with your data, until it runs. Then I hand over. Some companies need someone to own the AI decisions first.
Sound familiar?
Twenty questions I keep answering. One of them is probably yours.
A license does not build the system
The work is orchestration, evaluation and guardrails. None of that is included in the license.
Context management
Context engineering: what to load, what to summarize, what to throw away. Not a bigger context window.
Usually both
MCP for tools the agent picks itself. Plain APIs for the things it must never get wrong.
Self-hosted
Open models on your own hardware. I run a 26B one on a desktop CPU. No GPU, no vendor.
Agents, not chat
An AI agent takes the action and reports back. A chatbot only gives you text to read.
Agents that test their own work
AI coding agents that write the code, run the tests and fix what fails. Not a person pasting prompts into a chat.
Orchestrated
Multi-agent orchestration with backpressure, retries and crash recovery.
Evaluated
LLM-as-a-judge scoring against your own cases, re-run on every change.
Guardrails
Hard limits, required tests and prompt-injection detection before it is allowed to act.
Every input checked
Every input is checked for prompt injection before the agent reads it.
Checked
Every answer is checked against the source documents for made-up facts, and it says when it does not know.
Usually neither, yet
Most teams need better retrieval and real evaluation before either one is worth the cost.
1/100th
A fine-tuned small language model doing the large one’s job, at a fraction of the cost.
In production
Most pilots fail at orchestration and evaluation. That is the part I build first, inside your company.
It can
Agents connect to the old system through its existing interfaces. Legacy modernization without a rewrite.
Smaller than you think
Almost always. I’ll show you measured benchmark results rather than vendor claims.
1 hour
Agents run it continuously. People only handle the exceptions.
One workflow
Measured before and after, on a single workflow, before it is rolled out further.
Readable
Architecture the next team can understand, take over and keep running.
An honest no
I’ll tell you which half of this not to build. Knowing what not to build usually saves more than building it.
Anyone can generate code now. Using AI to find and solve real business problems at scale is the hard part.
,
contributions in the last year
Articles about the work.
What apeg.dev is.
apeg.dev is the website of Arun Prasath E G, an agentic AI architect and forward deployed engineer building AI systems inside companies worldwide.
Twenty years. Seven exits.
I take on one or two engagements at a time, inside the company, and I hand over when it runs. If you have a problem in your own processes that no single product solves, and you need it built and running, not just a recommendation, that is the kind of work I do.
Or find me here