How Beghou Slashed Configuration Time 10x With AI
10x
Not disclosed
Not disclosed
Beghou Arc instances are programmatically generated from 100+ interrelated metadata tables spanning 1,500+ columns. That flexibility meant even simple additions, like a search bar, required complex manual engineering across multiple tables and foreign keys, creating a scalability bottleneck as the client base grew.
Fractional AI built an AI-powered copilot for Beghou Arc that translates natural-language configuration requests into validated SQL proposals. Engineers keep full control by reviewing every change before execution, while the copilot handles the most repetitive work.
A FastAPI service integrated with Arc's UI runs a low-latency GPT-4-mini agent with a read-only query tool to prevent unintended modifications. Metadata about the current page, grid, and datasource is dynamically assembled as context, and a containerized environment using MSSQL snapshots enables non-destructive regression testing. Temperature tuning (1.0 to 0.4) and parallel tool calling improved accuracy and cut reasoning steps.
Best fit for software or platform teams whose highly configurable products require slow, manual database or metadata engineering for routine changes.
Up to 10x faster configuration
~50% fewer errors
Fewer reasoning steps
Chris Taylor
By adding an AI copilot to its Arc platform that turns natural-language configuration requests into validated SQL proposals. A GPT-4-mini agent assembles live metadata context and proposes changes that engineers review before running, cutting configuration from hours or days to minutes - up to 10x faster.
The copilot runs a low-latency GPT-4-mini agent behind a FastAPI service integrated with the Arc UI, with a read-only query tool against the MSSQL metadata database. XML-based system prompts, temperature tuning, and parallel tool calling improved accuracy.
Configuration tasks completed up to 10x faster - minutes instead of hours or days - error rates fell by roughly 50% after temperature tuning from 1.0 to 0.4, and parallel tool calling cut reasoning steps from 6-7 to 2-3 per request.
A specific build timeline was not disclosed. The work included model and temperature tuning and a containerized MSSQL-snapshot environment for non-destructive regression testing before rollout.
It is best suited to software or platform teams whose highly configurable products require slow, manual database or metadata engineering for routine changes.
Have a similar challenge?
