AI Agent: One Job, Evals and an Approval Step by Ergin KAI Agent: One Job, Evals and an Approval Step by Ergin K
AI Agent: One Job, Evals and an Approval StepErgin K
Cover image for AI Agent: One Job, Evals and an Approval Step
An agent that does one job inside your tools, and asks before anything it can't take back.

Who this is for

Teams with a repetitive job that needs judgement: reading inboxes, screening documents, researching accounts, proposing actions.

How I build it

The agent gets the tools the job needs and no others. Anything risky goes through a person: it proposes, they approve.
Failure is loud. A broken login is an error that reaches someone, never an empty result the agent reads as "nothing to do".
An evaluation set is written before the prompt, and every change has to pass it.
Tracing and a cost cap per account from day one.

Shipped examples

Caldra (a scheduling agent that proposes and never books on its own), Xandidate (screening that quotes the CV line behind every judgement), VC Automation (the model writes a plan, ordinary code runs it).
Fixed price and a written plan within three days of a free 30-minute call. The person on your first call writes your code.
Starting at$12,000
Duration5 weeks
Tags
Anthropic
Claude
LangChain
OpenAI
Python
Supabase
TypeScript
Vercel
AI Agent Developer
Service provided by
Ergin K proPristina
8
Followers
AI Agent: One Job, Evals and an Approval StepErgin K
Starting at$12,000
Duration5 weeks
Tags
Anthropic
Claude
LangChain
OpenAI
Python
Supabase
TypeScript
Vercel
AI Agent Developer
Cover image for AI Agent: One Job, Evals and an Approval Step
An agent that does one job inside your tools, and asks before anything it can't take back.

Who this is for

Teams with a repetitive job that needs judgement: reading inboxes, screening documents, researching accounts, proposing actions.

How I build it

The agent gets the tools the job needs and no others. Anything risky goes through a person: it proposes, they approve.
Failure is loud. A broken login is an error that reaches someone, never an empty result the agent reads as "nothing to do".
An evaluation set is written before the prompt, and every change has to pass it.
Tracing and a cost cap per account from day one.

Shipped examples

Caldra (a scheduling agent that proposes and never books on its own), Xandidate (screening that quotes the CV line behind every judgement), VC Automation (the model writes a plan, ordinary code runs it).
Fixed price and a written plan within three days of a free 30-minute call. The person on your first call writes your code.
$12,000