# Leon Liao on Building Checkie-AI as a Second Pair of Eyes, Not a Lie Detector > Source: > Published: 2026-08-12 09:11:29+00:00 Two things happened to Leon Liao on the same day. He helped a friend read a freelance contract that felt subtly wrong, and that evening he came across a dating profile that seemed a little too perfect. Neither had an obvious red flag. Both left the same feeling that something did not add up. By the end of the night he had a rough prototype open on his laptop. That prototype became Checkie-AI, an AI reality check tool built to help people notice inconsistencies and ask better questions before they commit to something. Liao, its founder, spoke to StartupFortune about why contracts and dating profiles turn out to be the same problem underneath, and why he is careful about what the product refuses to say. ## Where did Checkie start? Checkie really began with two unrelated situations that happened on the same day. Earlier that day I was helping a friend review a freelance contract because something about it felt off. Then later that evening I came across a dating profile that seemed almost too perfect. Neither one had a single obvious red flag. They just gave me that familiar feeling that something wasn't adding up. That was what stayed with me. In both situations, one person controlled how the story was presented, while the other had very few ways to question or verify it. The settings were completely different, but the imbalance was the same. By the end of the evening I had opened my laptop and started putting together a rough prototype. It wasn't polished, but the idea was already there. ## Contracts, freelance gigs and dating profiles are very different problems on the surface. What is the common thing underneath them? For me, the common issue is the distance between presentation and reality. People don't always lie outright. Often the warning signs are more subtle. Something important may be left out, a claim may be worded vaguely, or two details may sound reasonable on their own but stop making sense when you put them together. In a contract, that might be an unfavorable clause buried beneath an attractive offer. In a dating profile, it could be a lifestyle claim that doesn't match the rest of the person's story. Checkie looks at how all those pieces fit together: the wording, tone, context, individual claims, and the relationships between them. The goal isn't to stamp something as true or false. It's to show the user what may deserve a second look. That logic works across different use cases because the system is analyzing consistency and context, not just the type of content it has been given. ## Which use case did you build for first, and did people use it for something you had not planned? Contracts came first. I originally built the core engine to review freelance contracts and gig listings. I wanted to help independent workers catch misleading offers, potential scams and unfair terms before they committed to something. Dating profiles weren't part of the original plan at all. But people began pasting Tinder and Bumble bios into Checkie to see whether a potential match might be too good to be true. That caught me by surprise, but once I saw how people were using it, it made sense. It also changed how I thought about the product. I realized that AI profile verification could be useful well beyond professional situations. Questions about digital trust show up everywhere, from the contracts we sign to the people we choose to meet. ## What is the hardest part of telling someone their document or profile does not hold up? Finding a concern is only half the job. Communicating it responsibly is just as important. AI can sound very confident, even when it's working with limited information. I never wanted Checkie to act like a judge or hand down a verdict about someone it doesn't know. A potential client, business partner or romantic interest shouldn't be labeled a liar based on one document or profile. That's why I spent so much time refining the prompts, guardrails and overall experience. Instead of saying "this is fake", Checkie might say "these three details appear inconsistent, and you may want to clarify them before moving forward". That difference matters. A useful warning should create clarity, not panic. The aim is to help people ask better questions without making them feel embarrassed or telling them what decision they have to make. ## Plenty of tools promise to detect what is fake. Where should that kind of tool stop short? A lot of tools are built to check one particular signal, such as suspicious keywords or whether an image appears somewhere else online. Those checks can be useful, but they don't always tell you whether the overall story makes sense. Checkie asks a slightly different question: does this hold together? It examines context, intent, internal consistency and the relationships between different claims. Rather than treating one signal as proof, it brings several pieces together and shows the user where the story may need more scrutiny. At the same time, Checkie isn't a magic lie detector, and I don't think any AI tool should present itself that way. AI can point out blind spots, raise questions and help people notice potential risks. It shouldn't make the final decision for them. I think of Checkie as a second pair of eyes. It's there to support someone's judgment, not replace it. ## What was the hardest stretch of building it? One of the toughest periods was improving the processing speed without making the analysis less useful. The early versions were thorough, but they could be frustratingly slow, especially when someone uploaded a complex multi-page contract. I could see users leaving before their results were ready. At that point it didn't matter how capable the technology was. If people were stuck watching a loading screen, the product wasn't doing its job. I spent a few intense weeks rewriting the processing pipeline and improving how we handled API calls. It wasn't the most glamorous part of building the product, but it made a real difference. We eventually reduced latency by about 80 percent, and retention improved afterward. It reinforced a simple lesson: users don't experience your technical architecture, they experience the wait. ## Who is Checkie for right now, and what is next? Right now, Checkie is for people navigating digital interactions where the consequences matter. That could be a freelancer reviewing a contract, someone evaluating a dating profile, or a consumer trying to work out whether an online offer is legitimate. The situations may be different, but the underlying need is the same: people want more clarity before making a decision that could affect them personally or financially. Over the next year I'll be focused on expanding our API integrations. I want Checkie to work within the platforms where people already communicate, browse and do business. The best version of Checkie may not require users to open another tab or copy and paste anything. Ideally the support should appear when and where it's needed, with as little friction as possible. ## Where would you like Checkie to be in a few years? I'd like Checkie to feel less like a standalone destination and more like a quiet layer of support built into the internet. Spell-check is a good comparison. Most people don't think about opening a separate tool to check every sentence. It simply works in the background and catches mistakes before they send or publish something. I can imagine reality-checking working in a similar way, alerting people when a claim appears inconsistent, misleading or worth examining more closely. As generative AI makes convincing content easier to produce, we'll also need tools that help people slow down and evaluate what they're seeing. My long-term goal is for Checkie-AI to become part of that protective layer. It shouldn't tell people what to believe. It should show them why something may deserve another look and help them decide what to do next. If Checkie can make people feel a little safer and more confident in their online interactions, that would be a meaningful outcome.