Agent Builder

Problem

Every new client requirement at Mesha meant new bespoke endpoints, so the four-person engineering team became the bottleneck for work that followed patterns the product already had.

Solution

Designed an AI Agent Builder that lets clients assemble custom workflows from the company’s existing backend services on a plug-and-play basis. Every step in a workflow has human-in-the-loop review, so agents never act on client finances unchecked.

Architecture

Workflows are composed from existing service building blocks in a Next.js frontend and executed by the TypeScript/Express backend on AWS, with PostgreSQL and MongoDB for storage. Each workflow step pauses for human review before the next action runs.

Engineering decisions

• Composed new workflows from existing backend services instead of writing one-off integrations per client
• Human review at each step rather than autonomy at the end, so a bad output is caught before it compounds
• Building blocks shared across workflows so the second and third agent cost a fraction of the first

Technology

TypeScript · Express · Next.js · PostgreSQL · MongoDB · AWS

Results

• Clients could build custom agent workflows without waiting on new API development
• Cut development time for new client-specific workflows

What Daniel learned

Reliability came from structure, not model capability. Review checkpoints and reusable steps beat open-ended generation every time money was on the line.