---
title: "Discussion on \"Before AI Moves Money: The Eval Gate I Built for AutoKon\" | Hashnode"
description: "Discussion on \"Before AI Moves Money: The Eval Gate I Built for AutoKon\". Most AI features fail quietly. A summary is a little off, a label is wrong, a suggestion gets ignored. Nobody files a ticket and the product keeps working.\nAutoKon is not that kind of product. It is a"
url: "https://hashnode.com/posts/before-ai-moves-money-the-eval-gate-i-built-for-autokon/6ad4c7ca5d5bf6bbe9352fbf"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [Before AI Moves Money: The Eval Gate I Built for AutoKon](https://faishalwahiduddin.hashnode.dev/before-ai-moves-money-the-eval-gate-i-built-for-autokon?utm_source=hashnode&utm_medium=feed)

Most AI features fail quietly. A summary is a little off, a label is wrong, a suggestion gets ignored. Nobody files a ticket and the product keeps working. AutoKon is not that kind of product. It is a

Oct 18 · 5 min read · faishalwahiduddin.hashnode.dev  
Read the article

-   [#llm](https://hashnode.com/tag/llm)
-   [#artificial-intelligence](https://hashnode.com/tag/artificial-intelligence)
-   [#machine-learning](https://hashnode.com/tag/machine-learning)
-   [#fintech](https://hashnode.com/tag/fintech)

## Written by

[Faishal Wahiduddin](https://hashnode.com/@faishalwahiduddin)

@faishalwahiduddin · Bandung, Indonesia

  
[View profile](https://hashnode.com/@faishalwahiduddin)

Your tech talk partner — apps, websites & AI. Chief Technology Officer.

Chief Technology Officer based in Bandung. I built AutoKon's construction platform alone (2025 to 2026) and led engineering teams of 10 to 20+ at Wool and DoCheck. Fulbright M.S. CS (UCF). Site: https://faishalwahiduddin.com/?utm\_source=hashnode&amp;utm\_medium=profile&amp;utm\_campaign=blog-network-2026-10

## Responses 1

Markdown supported · ⌘↵ to post

[Yunus Emre Vurgun](https://hashnode.com/@yunusemrevurgun)

[Oct 18](https://hashnode.com/posts/before-ai-moves-money-the-eval-gate-i-built-for-autokon/6ad4c7ca5d5bf6bbe9352fbf/comment/6ad4d6c8b32f28f6765938af)

Putting the judge in the payment path and failing closed is the right call, and narrowing the prompt to per-claim grounding rather than overall quality is the key detail. One thing I'd be curious about: do you ever run the judge on a deliberately mismatched photo/description pair (e.g. a valid description attached to a photo of a different unit) as a canary? Since both models can share the same blind spots, a small rotating set of known-bad pairs is a cheap way to catch the judge drifting toward agreeing with the generator.
