A personal engineering lab by Bohea Still. Runnable prototypes, transparent post-mortems of where AI heuristics fail, and honest deep-dives into stubborn software problems. No marketing fluff. Nothing for sale. Bohea Still 的个人技术解构实验室。把晦涩、前沿的技术嚼烂吃透，讲给人听：能跑的代码实验，如实记下 AI 和算法在哪里会坏。非商业项目，无广告，不卖东西。
Why naive regex & LLMs destroy bank statements, and how deterministic balance checksums preserve columns. 为什么普通正则和大模型会把对账单借贷列搞崩？试用下面的双向校验解析逻辑。
Fully functional in your browser. Each includes a breakdown of its failure modes and architectural limits. 完全在你的浏览器本地运行，不上传任何字节。每个工具都附带失败案例与极限分析。
Splits plain-text statement lines into date, description, amount and balance, reads accounting negatives like (150.00), checks every row against the running balance and guards the CSV against formula injection. 把纯文本对账单行拆成日期、摘要、金额、余额四列；识别 (150.00) 这类会计负数，逐行用余额核对，导出 CSV 时防公式注入。
Removes page furniture such as "Page 3 of 12" and separator lines, turns column-aligned text into Markdown tables and shows how much shorter the result is. 删掉「Page 3 of 12」这类页眉页脚和分隔线，把按空格对齐的列变成 Markdown 表格，并如实显示长度变化。
Masks emails, US Social Security numbers, card numbers that pass the Luhn check, phone numbers and IPv4 addresses, entirely in your browser, before you paste text into a chatbot. 在浏览器本地抹掉邮箱、美国社保号、通过 Luhn 校验的卡号、电话号码和 IPv4 地址，粘贴给聊天机器人之前先脱敏。
Deep dives into hard engineering corners. Unfiltered conclusions backed by working benchmarks. 啃那些被包装过度、容易踩坑的硬技术。不抄新闻通稿，只讲实测出的硬结论。
Generic converters treat a statement as a picture of text, so signs, accounting parentheses and wrapped lines get lost. This parser keeps the sign separate from the number, reconciles every row against the running balance and flags the rows that do not add up.
Read the breakdown and run the code →Repeated page furniture costs context and gets quoted back as if it were content. A small four-step cleaner, what it removes, and the kinds of text it gets wrong.
Read the breakdown and run the code →A plain regex for long digit runs redacts every order number in sight. Checking the Luhn digit in the browser keeps those readable and masks only numbers that can be card numbers, without sending the text anywhere.
Read the breakdown and run the code →AI-written code, AI agents, and where software value goes. All notes → AI 写的代码、AI Agent，以及软件价值往哪里去。全部文章 →
My work is industrial software. At BoheaStill.com I build operator HMIs, device integration over Modbus, OPC UA and MQTT, and machine data into the customer's ERP, for systems integrators and machine builders. On a machine a bug can hurt someone, so that work is deliberately conservative: the most proven approach, verified before it ships, with safety resting on gates and hardware that my application cannot override. 我的主业是工业软件。在主站 BoheaStill.com，我给集成商和设备厂做操作员 HMI、通过 Modbus、OPC UA、MQTT 对接设备、把机器数据送进客户的 ERP。机器上的 bug 会伤人，所以这份工作刻意保守：用最成熟的做法，上线前先验证，安全交给我的程序改不动的闸门和硬件。
I am also curious about AI, agents and new software tools, and I like explaining them in plain language. Mixing those experiments into client machines would be wrong for both, so they live here. 我也对 AI、智能体和新的软件工具很好奇，喜欢把它们讲成大白话。把这些实验搅进客户的机器里，对两边都不对，所以它们放在这里。
One full industrial practice, plus a separate research habit. Not two half-time jobs. Client work always comes first; the lab is what I do with the rest of my time, and it keeps me honest about which new tools are ready for production and which are not. The lab is not a business: no subscriptions, no ads, nothing for sale. 一份完整的工业实践，外加一份额外的研究习惯，不是两份兼职各占一半。客户项目永远排在前面；实验室用的是剩下的时间，也让我清楚哪些新工具已经能上生产、哪些还不行。实验室不是一门生意：没有订阅，没有广告，不卖任何东西。
Found an edge case or want to discuss a hard engineering problem? Reach out directly: hi@byteplain.com 发现了解析边界 Bug，或者想交流硬核技术问题？直接给我发邮件：hi@byteplain.com