BytePlainlab
Discuss Tech 交流技术
Projects项目

Software work outside the factory 工厂之外的软件项目

Earlier client projects and capability demos from outside the factory: data extraction, payments, marketing operations, routing and AI pipelines. They are kept here for their engineering, not as a menu of what I build now, and each one comes with something you can check. My practice is industrial software, the machines and the software that ships with them, at boheastill.com. 工厂之外的早期客户项目和能力演示:数据抽取、支付、营销运营、路径规划、AI 流水线。放在这里是为了展示工程本身,不是现在的服务清单;每个都附带能自己核实的东西。我的主业是工业软件(机器和随机器出货的软件),在 boheastill.com。

Questions about any of them: hi@byteplain.com. 对哪一个有疑问,写信到 hi@byteplain.com。

Write-ups项目复盘

2026-07-15 · Automation, De-risked

A web form becomes a scored, routed lead 一个网页表单,自动变成已打分、已分派的线索

Enrich, score, store and notify: a lead pipeline proven end to end on mock data before any real credential is used. With public code you can run yourself. 补全、打分、入库、通知:一条在模拟数据上从头跑通的线索流程,用到真实凭据之前就已验证。附可以自己运行的公开代码。

2026-07-07 · Systems That Don’t Talk

Managing Google Ads campaigns in plain English 用一句自然语言,管理 Google Ads 广告

An MCP server that lets Claude pause campaigns, adjust budgets and manage keyword bids, with a mock mode you can demo without credentials. 让 Claude 直接暂停广告系列、调预算、改关键词出价的 MCP 服务器,带一个不需要凭据就能演示的模拟模式。

2026-07-07 · AI for Traditional Business

From chaotic backtracking to clean neighborhood sweeps 从混乱折返,到整齐的逐街区扫街

A vehicle-routing optimizer that finishes one district before opening the next, because real drivers don’t drive the way shortest-path math says. 车辆路径优化器:跑完一个片区再去下一个,因为真实的司机不会照着最短路径算法开车。

2026-07-07 · Systems That Don’t Talk

Doing what the Shopify API can’t, safely 安全地做 Shopify API 做不到的事

Retroactively linking historical orders to B2B Company Locations, a workflow Shopify’s API doesn’t expose, with automated 2FA and no password sharing. 把历史订单补关联到 B2B 公司地点,这是 Shopify API 不提供的操作;自动通过 2FA,不用共享密码。

2026-07-07 · AI for Traditional Business

A phone call becomes a spreadsheet row, automatically 一通电话,自动变成一行表格数据

Twilio → transcription → LLM extraction → Google Sheets, with a confidence score that decides when a human needs to look. Twilio → 语音转写 → LLM 提取 → Google Sheets,置信度评分决定何时才需要人看一眼。

Runnable repositories可运行的仓库

Each one runs without credentials and says what it does and does not prove. 每个都不需要凭据就能跑,并写明它能证明什么、不能证明什么。

AI Agents · GovernanceAI Agent · 治理

An agent that can propose a payment but never make one一个能提议付款、却永远付不了款的 agent

Run attack_it.py: fifteen attempts to move money that shouldn’t move, including prompt injection, self-approval, an amount swapped after approval, a replayed webhook, and an “anonymised” phone hash brute-forced in 4,568 tries. All fifteen fail in about a second, enforced in CI. Three of them were added after a reviewer broke an earlier version.运行 attack_it.py:十五次让不该动的钱动起来的尝试,包括提示注入、自己审批自己、审批后偷换金额、重放 webhook,以及一个“匿名”手机号哈希被 4568 次尝试暴力破解。十五次全部失败,大约一秒跑完,由 CI 强制执行。其中三项是一位评审攻破早期版本后补上的。

github.com/boheastill/llm-agent-payment-gates ↗
Infrastructure I run我自己在运维的基础设施

A fleet of MCP servers with no central registry一支没有中心注册表的 MCP 舰队

One MCP per directory, one port per MCP, discovered by scanning instead of from a list someone has to remember to update. Works across Windows and WSL, binds to loopback by default, enables auth with an environment variable instead of a code change, and documents the three root causes that actually took servers down in production.一个目录一个 MCP,一个 MCP 一个端口,靠扫描自动发现,不靠一份需要有人记得更新的清单。同时支持 Windows 和 WSL,默认只绑定本机回环地址,开启鉴权只需一个环境变量,不用改代码,还记录了生产环境里真正把服务搞挂的三个原因。

github.com/boheastill/hua-mcp ↗
AI & DataAI 与数据

Constrained ASR beats a general model in noise: 52% vs 14%受限词表 ASR 在噪声下胜过通用模型:52% vs 14%

Closed-vocabulary ASR for German number words: constrained decoding compared with a general model under noise, 52% vs 14% at 5 dB. Vosk and Sherpa-ONNX, running fully on-device.德语数字词的闭词表语音识别:在噪声下对比受限解码和通用模型,5 dB 时 52% 对 14%。Vosk 加 Sherpa-ONNX,完全在设备端运行。

github.com/boheastill/german-number-asr-demo ↗
AI & DataAI 与数据

Regulatory HTML into strictly validated JSON法规 HTML 变成严格校验的 JSON

Offline-first Python framework parsing government and regulatory HTML into strictly validated JSON. Proven on 233-article statutes with zero warnings.离线优先的 Python 框架,把政府/法规 HTML 解析成严格校验的 JSON。在 233 条条文的法规上验证,0 警告。

github.com/boheastill/gov-doc-parser-framework ↗
AI & DataAI 与数据

Floor-plan PDFs into structured data建筑图纸 PDF 变成结构化数据

Architectural floor-plan PDF → structured JSON/Excel (room stamps, areas, fixtures), with a low-confidence validation layer that flags what a human should check. Layout-aware, German-field sample.建筑平面图 PDF → 结构化 JSON/Excel(房间标注、面积、设施),带低置信度校验层,标出需要人工复核的部分。版面感知,德语字段样例。

github.com/boheastill/floorplan-pdf-extractor ↗
AI & DataAI 与数据

Customer segmentation that survives validation经得起验证的客户分群

Segmentation → personas with validation beyond statistical separation: bootstrap stability, holdout generalization, an external-outcome check, and a 3-gate k selection (silhouette / stability / distinctness).分群 → 画像,验证不止于统计可分性:bootstrap 稳定性、留出集泛化、外部结果校验,以及三闸门 k 选择(轮廓系数/稳定性/区分度)。

github.com/boheastill/b2b-segmentation-demo ↗
Systems & Integration系统与集成

A four-tier embedded payments architecture, simulated四层嵌入式支付架构,可交互模拟

Multi-tier embedded payments (league → team → player → parent): a technical proposal plus an interactive simulator showing how transactions are tokenized, routed and split across accounts.多层嵌入式支付(联盟 → 球队 → 球员 → 家长):技术提案 + 交互模拟器,演示交易如何被令牌化、路由并在账户间拆分。

github.com/boheastill/embedded-payments-hierarchy ↗
Systems & Integration系统与集成

A private, self-hosted stream between your devices and AI设备与 AI 之间的私有自托管通道

A self-hosted, database-free chat stream for moving text and files between PC, phone and AI agents. A single Go binary with token auth, built privacy-first. It’s the worked example behind my note on the outbound boundary.自托管、不用数据库的聊天流,在电脑、手机和 AI agent 之间传文字和文件。一个 Go 二进制文件,token 鉴权,优先保护隐私。也是我那篇“出站边界”笔记里用到的实例。

github.com/boheastill/Intranet-Chat-Stream ↗
Automation自动化

Bounce-rate monitoring with a credential-free mock API退信率监控,自带免凭据 mock API

Monitors Klaviyo campaign bounce rates and alerts to Slack: an n8n workflow plus a Python monitor, with a mock Klaviyo API so the whole thing can be demoed without credentials.监控 Klaviyo 营销邮件的退信率,超标就发 Slack 告警:n8n 工作流加 Python 监控程序,自带模拟的 Klaviyo API,整套不需要凭据就能演示。

github.com/boheastill/klaviyo-bounce-monitor ↗
Automation · Live自动化 · 在线可玩

Route optimization you can pan and zoom, live可以拖动缩放的路径优化,在线

Fleet routing with OR-Tools CP-SAT: 20% less mileage on identical stops through district-contiguity constraints. The interactive map is live, so you can open it and follow a single vehicle through its district.用 OR-Tools CP-SAT 做车队路径规划:同样的站点,靠片区连续性约束少跑 20% 里程。交互地图在线可用,打开就能跟着某一辆车走完它负责的片区。

boheastill.github.io/montreal-district-vrp-solver/vrp_map_montreal.html ↗
Performance & Backend性能与后端

1,000 pages: 50 minutes → 2.5 minutes1000 页:50 分钟 → 2.5 分钟

Fast PDF-to-image rendering (PDFBox) with a custom bounded thread pool, taking roughly 1,000 pages from 50 minutes down to 2.5.用 PDFBox 快速把 PDF 渲染成图片,配一个自定义的有界线程池,约 1000 页从 50 分钟缩短到 2.5 分钟。

github.com/boheastill/pdfSplit ↗
Performance & Backend性能与后端

A Django slice you can read一段可以读的 Django 切片

Django 6 slice: a pseudonymous community forum plus a bookable-advisor marketplace. Explicitly a code sample to read, not a claim.Django 6 切片:匿名社区论坛 + 可预约顾问市场。明说了:这是一份供阅读的代码样例,不是一个声称。

github.com/boheastill/mainstep-community-demo ↗