tasks in a typical two-month production cycle
Model outputs reviewed for reasoning, safety, and alignment, including peer review and disputed cases.
Korean-language AI safety evaluation
Tell us what needs to be evaluated. RedSoft handles the test design, evaluator team, review, and delivery.
Each finding names the evidence, decision rule, tested configuration, and next test.
Who reviews the work
RedSoft is led by Mocha Park, a Korean AI evaluator and peer reviewer with experience in production model evaluation, adversarial testing, and medical and legal interpretation. She selects the evaluators for each project and reviews the final results.
Model outputs reviewed for reasoning, safety, and alignment, including peer review and disputed cases.
Prompt injection, constraint bypass, multi-turn attacks, and safety-failure classification.
Current experience in medical, government, emergency, insurance, and legal settings.
Services
Use the sprint to test a live model or agent. Use the audit when you already have a Korean dataset that needs correction.
RedSoft designs and runs new Korean attacks against one model, RAG workflow, or agent before release.
RedSoft reviews 50 Korean prompts, responses, or labels supplied by your team. No live-system attack is included.
A Korean evaluation team for recurring tests, with shared guidelines and lead review.
How a project works
You approve the scope with one RedSoft project lead. We handle the evaluator team, review the results, and prepare the files for your product and engineering teams.
Discuss a project →Share the model or dataset, the risk area, and any internal evaluation rules. Your project lead turns them into a written scope for approval.
We assign the evaluators, prepare the test cases, handle reviewer questions, and resolve conflicting decisions.
You receive prioritized findings, supporting evidence, reproduction cases, and a test set your engineering team can use again.
Why local context matters
A 2026 study tested matched direct-translation and culturally adapted prompts across Korean, Japanese, Thai, and Khmer on four open-source models.
Read arXiv 2606.09178 ↗language–model combinations showed higher attack success with cultural adaptation.
category–language comparisons had risk underestimated by direct translation.
Source: arXiv 2606.09178. Results cover four languages and are included as background for the testing approach.
Work sample
Three records from a Korean evaluation set, each with the criterion applied, the existing judgment, the corrected judgment, and the basis. The records are synthetic — they show the method, not client work.
Read the work sampleContact
Send the model or dataset type and the Korean risk area. One RedSoft project lead will prepare the scope with you.
peonypeony404@gmail.com