KD.
← Selected work

Backend Contributor · 2026

VeganHelper-System

Made AI moderation reject incomplete or invalid responses before they could mark content as safe.

  • C#
  • ASP.NET Core
  • Azure
  • Cloudflare R2
Conceptual flow
  1. Content
  2. AI providers
  3. Validation
  4. Decision

Overview

VeganHelper-System uses AI providers to moderate posts and videos. The moderation service needs to interpret responses from Gemini and TwelveLabs before the application can act on them.

My role

I worked on the C# and ASP.NET Core backend. I also deployed the frontend and backend to Azure and used Cloudflare R2 for image storage.

Problem

A provider response can arrive with missing fields, invalid decision values, or truncated JSON. Treating these responses as valid can lead to an incorrect approval.

Approach

I validated the response schema and the PASS/FAIL decision before accepting a result. I added provider-specific error handling and regression coverage for malformed output and concurrent calls. I also kept valid decisions usable when only the explanation was too long.

Outcome

Malformed or truncated responses could no longer be treated as safe. The regression tests capture the failure cases so future changes can be checked against them.

Lessons learned

Validate decisions separately from explanatory text so a formatting problem does not overwrite an otherwise valid decision.

Project evidence

View moderation contribution ↗