Solo Dev + AI, Building an Enterprise System: Getting MakanCloud's HR Module Ready to Ship

This week I shipped MakanCloud v8.34.0. The main work in this release is stage S5.8 of the HR module.

MakanCloud is the first commercial product I've built since quitting my job. It's an end-to-end system for F&B businesses in Singapore. It covers ordering, the central kitchen, delivery, staff scheduling, POS and QR-code ordering. I do all the development and ops myself, with AI helping.

People often ask me: can one person with AI really build an enterprise-grade system? I don't want to get philosophical about it. Instead, here's what I actually did this week.

1. Settle the rules first, then write code

This stage had 7 business questions that needed a final decision. Internally I numbered them D12–D15 and D17–D19. Some examples: which data has to live in a private directory, what the dashboard shows to each role, and how to tell actions taken by the platform apart from actions taken by the store itself.

I write every decision down in a response checklist and a principles doc first. Only after the rules are settled do I hand the work to AI. This feels slow, but it saves the most time. If you let AI write code before the rules are clear, it'll write fast. It'll also head in the wrong direction just as fast.

2. Don't rush to ship: do a security review first

Before merging the feature, I did a dedicated security review, and it found real problems:

• An old data channel hadn't been fully shut off, so the new rules could have been bypassed.
• The dashboard API's "viewer role" used to be an optional parameter. If it was left out, the permission check ended up in a gray area. It's required now.
• I added a source flag to the "is this a platform operator?" check, so actions taken on the platform side don't get logged as if a store employee did them.

If these three had gone to production, you probably wouldn't have noticed them day to day. But they're exactly the permission and audit issues enterprise customers care about most.

3. Data should be traceable, deploys should be reversible

I added per-batch auditing to the opening-data import script. For every batch there's a record of when it was imported and who imported it.

Before deploying, I took a snapshot of the server. I only let the AI run the deploy after I'd confirmed the snapshot was done. I also moved the sensitive private directories onto a separate encrypted attached disk. I scheduled that migration for after the stores closed for the night so it wouldn't get in the way of daytime orders.

After the deploy, I updated the status doc, the README, the handover notes and the deploy log. When you're building solo, your docs are really written for yourself three months from now.

4. What I actually think about "one person + AI"

AI really has made me write code many times faster. This week alone I had a dozen-plus tasks moving on MakanCloud. But this week also made me more sure of one thing: AI can handle the execution, but the judgment calls are still on you.

The bar for an enterprise system isn't how much code you write. It's in these places: have you really thought through the business rules, are there holes in your permissions, can your data be traced, and can you roll back when something breaks. I didn't skip a single one of these steps. AI just gives me, as one person, enough energy to get through all of them.

That's also why I'm confident taking on custom software work now. I build ops systems, internal tools and data dashboards for F&B businesses and SMEs, and I use the same process I use for MakanCloud. I own everything myself, from requirements to launch to ongoing ops.

If you need something similar, or you're also building with AI and have hit the same kinds of problems, leave a comment or DM me. I'd love to chat.