# Yolanda Kim, VP of Platform Safety & Compliance at Threadwell — read of AI Image Moderator, June 5 2026 > Nine years in trust and safety, currently managing a two-person moderation team plus an offshore vendor for a B2B community platform with 80k MAU split across South Korea, Singapore, and the US. I ride the 7:15 BART every morning and listen to Korean true-crime podcasts. My daughter is four and already more opinionated about content standards than our legal team. --- ## How I got here We are in a mild panic about KOSA adjacent compliance timelines and our current stack (AWS Rekognition plus a Zendesk-based manual queue) is held together with prayer and a very stressed contractor in Manila. I Googled "image moderation tool South Korea MDA compliance" last Tuesday. This page came up third. I clicked because the meta description said "jurisdiction-aware" and that phrase is rare enough to be interesting. ## What I clicked first "Jurisdiction-aware detection" in the hero. That phrase is doing a lot of work and it is the right phrase. Most tools I have evaluated in the last six months say "global" which means "we trained on US data and good luck." The fact that they called out South Korea and Singapore by name in the first three lines made me keep reading. Also "Full audit trails for regulators" is the right second beat. Whoever wrote this knows that the pain is not the moderation, it is the quarterly audit. ## Where I paused The 0.8 second scan time claim and then immediately "99.7% accuracy on trained content categories." I stopped here for maybe two minutes. Both numbers are specific enough to feel researched rather than invented, which is more than I can say for every competitor I have looked at. But "trained content categories" is a weasel qualifier. Accuracy at WHAT? Explicit imagery? Violence? Political content that is legal in the US and not in Singapore? The claim is almost there and then it hedges. I wanted a footnote. There is no footnote. ## What I distrusted Two things, and the second one is a dealbreaker. First: "reduces team overhead by 70%." This is the kind of number that gets printed on a slide and then quietly disappears when you ask for the methodology. 70% of whose team? A team of two? A team of fifty? Based on what time period? This is not a claim, it is a feeling dressed up as a statistic. Second, and this is the part that completely reframed the page for me: I scrolled to the bottom and found this: "Honest disclosure: we don't have live customers on this idea yet. We shipped the strategy package; you ship the customer conversations." So this is not a product. This is a pitch deck wrapped in a product-looking website. The "Start Free Trial" button at the top is for something that does not exist yet. I genuinely did not know this until I hit the footer. The Starter plan at $299/month, the FAQ about GDPR compliance, the "standard platforms are live in 24 hours" -- all of it is hypothetical. I would have forwarded this to my CTO before I noticed. The "63/100 Adoptability" score and "$-23,120 Year-1 take-home" section is actually the most honest writing on the page, but it requires reading all the way to the bottom of a page that presented itself as a real vendor. ## What would convince me If this were a real product: a compliance attorney's letter confirming the jurisdictional ruleset was reviewed by someone with actual regulatory expertise in KR and SG. Not a blog post. A named attorney or firm. Regulated markets have specific, sometimes monthly-updating, content standards. "We update regional models quarterly" is not reassuring if a regulation changes in January and your next update is April. If this is a "buy the idea and build it" product: I want to see the actual regulatory citations behind the jurisdiction claims. Which law in South Korea? Which MAS guidance in Singapore? The strategy package is only useful if the underlying research is sound. ## What I'd ask in an email reply 1. The FAQ says images are processed in-memory and never stored. How does your audit trail work then? You cannot have both "we don't store images" and "full timestamped logs of every scan" without explaining the architecture in more detail. Regulators want to see the image alongside the decision, not just a hash. 2. South Korea's content standards are enforced by the KCSC and they are aggressive about what counts as "harmful to youth." Has anyone on your team actually corresponded with them, or is this model trained on publicly available takedown notices? 3. Mainland China is listed as a supported jurisdiction. That is not a moderation problem, that is a legal and political problem that most US-incorporated companies quietly remove from their feature lists after talking to a lawyer. What is the actual scope of support there? ## Verdict: dismissive Not because the idea is bad. The idea is genuinely good and the pain is real. But this page presents as a working SaaS vendor and is actually an idea package being sold for $99. A compliance buyer in a regulated market cannot evaluate a vendor that has no live customers, and the disclosure is buried below the fold after three rounds of pricing tables and a detailed FAQ. That sequencing is a trust problem. Fix the framing and this could be the first one of these I forwarded to procurement. --- *Memo by skeptic persona, generated 2026-06-05. Studio breaks own self-grading loop.*