Original Summary

What it does You upload the assignment brief your professor handed out — usually a 20-40 page PDF of requirements. The system reads it and produces the whole deliverable package: a formatted Word document, a slide deck, a defense speech, and, for engineering courses, A1 technical drawings in DXF/DWG with a GOST 2.104 title block. The student does the final pass and owns the result. Why it is not a ChatGPT wrapper Generation turned out to be the easy part. The year went into the layer that stops the model from shipping garbage. Production runs as a three-wave parallel agent swarm. Wave 1: one sub-agent per chapter doing research and calculations, returning structured JSON. Wave 2: a coordinator de-duplicates figures across chapters, then one agent per figure. Wave 3: parallel section writers. The coordinator writes the intro, conclusions and bibliography itself. Every figure must carry a justification AND evidence — a quoted page from the brief, a sample id, or the filename of the computed results. Producing fewer figures than planned is a hard failure, not a warning. 12 automated validators run between stages: citation coverage, block-formula preservation, broken figure references, numeric consistency, physical-magnitude sanity, work-type matrix, subsection policy and more. They aggregate every complaint into ONE rejection instead of round-tripping the model seven times. Stage flags are sha256-signed. The model cannot quietly raise its own page limit to sneak an oversized document through. A builder-marker check stops it from swapping in a hand-rolled script instead of the official document builder. Two engines I had to write from scratch Presentation engine: a constraint solver over 18 layouts and 12 themes, WCAG-validated colour, seed-based variation. I got tired of every AI deck looking like the same template. CAD engine: text to DXF/DWG. 18 vectoriser iterations before it was usable, plus a GOST 2.104 title block generated in code rather than pasted in as an image. Humanizer: 150+ patterns across 6 languages, wired into the validator so formulas and variable names cannot drift while the text is being rewritten. The numbers, including the bad ones ~2,960 unique visitors a month, 83-248 a day, fully organic, zero ad spend. Largest country is the US. 33 completed orders on disk, 867 MB of delivered work. ~55,000 lines of core source, about 23 TODO/FIXME markers in all of it. Revenue: $0. The wall I hit Five payment integrations are written and tested (Stripe, Paddle, CryptoBot, LiqPay, Fondy) and a successful payment auto-triggers the AI run. But I never connected live merchant keys, and order flow stopped in June when I moved to other work. So I have a system people find on their own and a checkout that cannot take their money. Also 4 of 15 variation knobs in the presentation engine are still unwired — documented, roughly 3-5 days of work. I am at the point where the honest move is to hand it to someone who will actually run it, so it is listed

中文概览

中文标题： 作者出售自动生成大学作业的AI系统

作者花一年构建AI系统，可上传教授的作业要求PDF后生成Word文档、幻灯片、答辩讲稿及工程课程的DXF/DWG技术图纸。系统采用三波并行代理架构、12个自动验证器和自研演示与CAD引擎。每月约2960自然访客，完成33笔订单，但收入为零，因未接入商户支付密钥。作者决定出售该系统。

  • 情报分类：开源项目与落地
  • 分类依据：作者出售自建AI作业生成系统，属项目展示
  • 信息来源：Reddit · SideProject
  • 发布时间：2026/9/13 09:53:40