Case study: I spent ~$50 to build a full-fledged SAAS with DS, and here what I learn from it

Case study: I spent ~$50 to build a full-fledged SAAS with DS, and here what I learn from it 图片 1

preview.redd.it/50o0wr52f5fh1.png preview.redd.it/46dgp2w3f5fh1.png I started my journey in the last week of May 2026. My background is software engineer & enterprise architect, and I've been building software (mostly SAAS) for the last 20 years. I occasionally do enjoy putting my money in some eSports game betting, most of the time I lose. So with my knowledge, I decided to start a SAAS application, first and foremost for myself to use. At the beginning, I used Opencode Go & Zen to test out several models (GLM, DS, MiniMax, etc.), along with Claude Code (from the 7-days pass) and Gemini (with Code Assist, and later AntiGravity). After a week since the beginning, I decided to stick with DeepSeek, but I also used some other models (Opus & Gemini), for reasons I will mention below. At the moment, I spent about 3b tokens in DeepSeek platform, at the cost of $30; plus $15 I spent on Opencode platform. I used CodeAssist, and now Antigravity on free plan. About the SAAS: It's an event-driven, not truly microservices. At the moment, I have 20 dockerized services running, spread across 2 VPS. API is built with ASP.NET, frontend is vite, ML/AI services + scrapers are built with python, db is postgresql. The application is heavily tested, from unit tests, backtest to e2e tests. I'm running a multi-models for predictions. At the moment, I'm using XGBoost and LightGBM, with a consensus mechanism for multi-models to produce predictions. I'm planning on adding a third model. I heavily utilize Cloudflare, from DNS, CDN, edge caching, to even SSR with workers (for SEO). This is 1-man project with no-one else beside myself involves. Most of the operations are automated or AI-assisted (using Deepseek v4-flash). After a lot of trial & errors, this is what works for me: I have all projects in the same directory (or folder). Then I use opencode on that said directory. I maintain an always-up-to-date knowledge-base. All information, from the beginning of the project, is structured in .md files. Every decisions that I made, every mistakes are in all those .md files. I also maintain agents.md file, which auto-generated at the beginning, but I heavily modified it. It contains rules: such as use gh cli to watch the CI action, or never ask for SSH credentials and always wait for my instruction to ssh. Always plan first, execute later. I use oh-my-openagent, specifically Prometheus to build a plan, which will spawn librarian & oracle agents; then I use Momus + Metis to detect gaps, verify the plan, but not on all plans, only logic-heavy plans; after that Atlas to execute the plan, which in turn will spawn several sisyphus-junior agents. After the execution of each plan, I will ask opencode to update the knowledge base, put in all decisions/learnings. I don't always put myself in the driver seat, but I see myself as a navigator/decision maker. Oracle is awesome for researching topics or assisting me to make decisions. My oh-my-openagent config my \"work station\" What doesn't work for me, and solution: Deepseek v4, both pro and flash, are terrible at UI/UX designing: At the beginning, I need to have a design system, but DS v4 was just terrible at it. I found that GLM, Opus, and especially Fable did an amazing job at this. I had those 3 models generate different design system for me, and at the end I chose the one that Fable generated. Deepseek v4-flash is not good at reasoning (duh). Solution is simple: always plan first, plan to the smallest details. Then use Deepseek v4-flash to execute (in my case, I use sisyphus-junior to execute). Deepseek v4-pro is really good at reasoning, but... sometimes it falls flat. Eg: I had DS4-pro looked into some features engineering issues, it catched 80% of the issues, but the rest 20% were really important. Opus was able to catch 85%, Fable was able to catch around 92-95%. Final words: from my experience, Deepseek v4 is just amazing but it's not an all-rounder, but you must know its capabilities. Use it the correct way and you will have amazing experience. PS: The product I built is zest.win if anyone is interested. I'm giving 1 month free access for any new registration.

添加评论
点赞收藏
点踩分享查看原文
评论
?
参与讨论