We built a complaints and case system for an employment council. Staff log each labour complaint once, an AI model suggests its category and a likely resolution, and designated agents see the case through to a final judgement. The client is kept anonymous on this page.
01 — The problem
Labour complaints reached the council on paper and by phone. There was no one place to see who was handling a case, what stage it had reached, or how long cases were taking.
Complaints also come in mixed. Unpaid wages and a dismissal can arrive in the same letter, and someone has to read each one and decide where it goes. The council needed:
02 — How we built it
Here's what each step looked like on this project.
Every complaint gets one record and a case number. Two kinds of user: admins, who see everything and assign cases, and designated agents, who work the cases assigned to them. The AI may suggest, but a person always decides.
We fixed the case stages (Pending, Assigned, Investigating, Resolved or Dismissed), kept the AI as a separate small service, and decided that every action on a case is logged. These choices are explained below.
The screens staff use most (the dashboard, logging a new complaint, a single case and the reports) are mapped as a clickable prototype with sample data, so you can follow a complaint through the system. Try it below.
The pieces came one at a time: complaint intake with document uploads, the case list with filters, assigning cases with an email to the agent, the AI suggestions, a timeline for each case, printable case reports, agent performance figures and a public status check.
We trained and checked the model in notebooks on labelled sample complaints. In the app, uploads are limited in size and file type, pages check the user's role, and if the AI service doesn't answer within a few seconds the complaint is still saved.
The system runs as a PHP and MySQL web app with the Python model running beside it. It includes a built-in user manual, and admins can export resolved cases as training data for the next update of the model.
03 — Try it
This is the clickable prototype, running live on the page. Every name, case and figure in it is sample data, and it may be a step behind the real system.
04 — Key decisions
In plain English, without the jargon.
When a complaint is logged, the model suggests one of five categories (such as wage dispute or unfair dismissal) and a likely resolution. The designated agent investigates and records the final judgement, which is what counts.
Each suggestion comes with a confidence percentage and the second most likely category. Complaints often mix issues, like unpaid wages and a dismissal, so staff see when the model is unsure instead of trusting a single label.
The model is a small separate service on the same machine. The case system asks it for a suggestion when a complaint is submitted. If it doesn't reply within a few seconds, the case is saved anyway and marked "Not predicted", so no complaint is lost.
Opening, assigning, updating and resolving a case are all written to that case's timeline: who did what, and when. The printable case report includes the timeline, and complainants can check their status with just their case number.
Urgency is deliberately simple. Words like "harassment" or "threat" mark a complaint High, and words about unpaid or late wages mark it Medium. Rules like these are easy for staff to understand. The dashboard also flags any case that has sat untouched for 30 days or more.
05 — Tech stack
PHP 7.4+, plain HTML, CSS and JavaScript, Chart.js for dashboard charts
MySQL for cases, documents, users and the case audit log
Python and scikit-learn: TF-IDF text features with logistic regression, able to give several categories per complaint
A small Flask API that the PHP app calls when a complaint is submitted
PHPMailer, so agents get an email when a case is assigned to them
Apache, PHP and MySQL (XAMPP), with the model service running alongside
06 — What the client got
Alongside case intake and tracking, the council got:
Each resolved case, with the agent's final judgement, can be exported as training data, so the model can be retrained on the council's own cases.
Your project
We follow the same steps for every client, and you see a clickable prototype before you pay for the build.