These are client projects, so the code is private. Here is what they do.

A complete internal platform for a fulfilment company. It brings seller operations, carrier investigations, customer support, warehouse attendance, credits and analytics into one place. Before it existed, the team worked out of a shared mailbox, a pile of spreadsheets and a whiteboard on the warehouse wall.

16 product pages, 97 backend routes, 5 workspaces, role aware access.

flowchart TB
    A[Gmail channel] --> D[Unified queue]
    B[Platform tickets] --> D
    C[Seller portal] --> D
    D --> E{Routing and<br/>ownership lookup}
    E --> F[Case record]
    F --> G[Carrier email<br/>cancellable send]
    F --> H[Credit registration<br/>internal ledger]
    F --> I[Task assignment]
    G -.carrier reply.-> D
    F --> J[Analytics<br/>and CSV export]
    F --> K[Audit trail<br/>who did what, when]

    style E fill:#1f6feb,stroke:#1f6feb,color:#fff
    style G fill:#d29922,stroke:#d29922,color:#fff
    style H fill:#d29922,stroke:#d29922,color:#fff

A few decisions worth explaining. The client's fulfilment platform stays read only: credits write to an internal record and never change anything on their live account. Anything that moves money or changes state can be undone, and carrier emails go out through a send window you can still cancel. Analytics runs on Amsterdam day boundaries, because a working day for this client is a Dutch day and not UTC. Get that wrong and every daily number is quietly incorrect.

Built for a logistics operator in the Netherlands. When a parcel went missing or turned up damaged, someone had to collect the order details, work out which carrier to contact, write the email, attach the evidence, and then chase the reply somewhere else. About 90% of that manual email work is now gone, and the system runs in daily operations.

The portal lists shipped orders that can still be investigated, then walks the seller through a two step form. Validation rules change depending on the carrier, since they each ask for different things. Evidence files get uploaded in the same flow. The system writes the carrier email, sends it to the right address, and pulls carrier replies back into the same case thread instead of into someone's inbox. The interface is in Dutch, because the operations team is Dutch.

Made for a screen on the warehouse wall rather than a desk. It shows who is working in which zone right now (pick and pack, replenishment, inbound, containers, returns), which tasks are open, and which ones are urgent. It refreshes every 10 seconds, so a supervisor can spot a bottleneck without walking over to ask.

It also produces 9 different CSV and Excel exports for managers, including a multi sheet workbook with daily trends and a summary for each department.

Most support bots say "I don't know" and the customer gives up. This one answers only from the company's own documents, checks how confident it is, and when it is not sure, it pulls a human agent into the same chat. The customer never has to start again or repeat themselves.

There are three parts: a chat widget you drop into any website, a backend that does the retrieval, and a live agent dashboard. Questions the bot could not answer go into their own queue, so the team can see exactly where their documentation is thin.

This one turns zoning PDFs into costed feasibility reports. It reads the documents and pulls out the constraints: plot area, setbacks, coverage limits, site conditions. Every value it extracts comes with the sentence it was taken from and a confidence score, and a person has to approve or reject it before it goes anywhere near the cost calculation.

Approved constraints then feed into scored design concepts and a cost estimate that lists what it does not know, rather than hiding the gaps.

Most of my projects follow the same path.

flowchart LR
    A[Manual workflow<br/>done by hand] --> B[Structured<br/>intake]
    B --> C{AI or logic<br/>layer}
    C -->|confident| D[Automated<br/>action]
    C -->|not sure| E[Human<br/>checkpoint]
    E --> D
    D --> F[Dashboard<br/>and exports]
    D --> G[Alerts and<br/>audit trail]

    style C fill:#1f6feb,stroke:#1f6feb,color:#fff
    style E fill:#d29922,stroke:#d29922,color:#fff
    style A fill:#21262d,stroke:#484f58,color:#fff

The yellow step is the one most people skip. Confidence checks, review steps, undo buttons, and a record of who did what. In a real business a wrong answer delivered confidently costs more than a slow one, so I build systems that stop and ask instead of guessing.

Automate the routine. Clarify the exception. Keep the human in control.