reddit · SideProject · 09-28 20:50 54 分 ← 返回列表

Laya/JEV play Magic the Gathering

Riding the JEV hype, I decided to build a project using the tech. Actually used Laya, the open-source version. The idea was to use the model to help with Magic: The Gathering Draft. For anyone unfamiliar, in Draft you pick a card from a pack, pass the rest to the next player, and end up building a 40-card deck. Fine-tuned 2 different models. The first picks the best card from the pack, considering previous picks. The second takes the cards picked during the Draft and helps bui

五维评分

付费意愿

需求真实度

个人可落地

竞争空间

新鲜度

AI 分析

痛点：MTG Draft 玩家选牌和组 40 张牌时缺乏实时、个性化的 AI 辅助，容易因选牌水平差而输掉比赛。

目标用户：Magic: The Gathering / MTG Arena 的 Draft 玩家，尤其是胜率一般、希望提升选牌和组牌水平的中轻度玩家。

证据：since I'm pretty bad at picking cards, it helped a lot — finished 5-3 and 4-3. / matched about 70% of players' actual picks and built decks with an expected win rate of at least 60%.

变现：Freemium 订阅：免费提供基础选牌建议，付费解锁实时 Draft 覆盖层、无限次分析、历史复盘、胜率预测和个性化教练；也可做 $5-10/月 Pro 版或按赛季订阅。

MVP：2 周内做一个轻量 Web/浏览器扩展 MVP：用户手动或半自动输入当前牌包和已选牌，调用已微调模型返回推荐选牌与前 3 理由；Draft 结束后根据已选牌推荐 40 张套牌，并用简化胜率模型给出评分。

竞品：17Lands、Draftsim、Arena Tutor、MTGA Assistant、Untapped.gg

摘要：作者用 Laya/JEV 微调了两个 MTG Draft 模型，一个选牌、一个组 40 张牌，训练数据来自 17Lands 中胜率 58%+ 且胜场 100+ 的玩家。测试中选牌匹配约 70% 玩家实际选择，组出的套牌预期胜率至少 60%，本人实测拿到 5-3 和 4-3。项目已开源到 GitHub 和 Hugging Face，但原帖没有付费意愿或收入数据。