Overview

From March to July 2024 I worked as an Enterprise Associate at Birmingham City University on an Innovate UK funded Accelerated Knowledge Transfer (AKT2I) project with PBL Care Limited, a home care provider. I led the feasibility study for automating PBL Care's complaint handling and built a proof of concept, which we called EchoCare.

The problem

PBL Care handled complaints, feedback, and operational requests on paper. Complaints came in by phone, in person, by letter, and by email, and were processed by hand. That created four problems:

• Delays and errors: manual processing slowed responses and increased the chance of human error
• Hard to analyse: paper archives made it hard to spot trends or serious incidents
• Long resolution times: the procedure allowed up to 28 working days to resolve a complaint
• Hard to access: the process was difficult for service users with impairments or limited technical skills

All of this had to work within the Care Quality Commission (CQC) standards for service user engagement, responsiveness, safety and safeguarding, and equity.

What I did

• Mapped PBL Care's six-stage complaint life cycle, from receipt and acknowledgement (within 3 working days) through investigation, resolution, escalation, and record keeping
• Catalogued the ten internal and external forms the business relied on, from spot checks to probation reviews, with the complaint types each one captures and who should be able to see it
• Reviewed CQC, GDPR, and Caldicott requirements, then built privacy by design, data minimisation, and role-based access into the design
• Carried out data protection impact considerations and documented the risks and retention needs
• Benchmarked ten existing tools, including osTicket, OTRS, Zammad, Zendesk, Freshdesk, and Salesforce Health Cloud, against cost, compliance, AI/ML capability, security, and scalability
• Designed and built the proof of concept described below

The proof of concept

• Digital forms: CQC-standard web forms for feedback, complaint investigations, and spot checks, replacing the paper versions. Each submission gets a unique ID and is categorised as a complaint, compliment, or suggestion.
• Database: submissions are stored in a MySQL database, with a separate store for internal forms and digitised historical records.
• NLP classification: Python (NLTK and spaCy) classifies complaints by type and tracks keyword frequency to surface the most common issues.
• Dashboard: a Chart.js dashboard shows complaint investigations, spot checks, feedback volumes, and sentiment in real time.
• API integration: the system connects through APIs to PBL Care's scheduling and planning application, feeding a single database and a unified dashboard.
• Data validation: a lightweight Python/Node.js GUI tool checks data validity and integrity. It was later merged into PHP for the API integration.

Research findings

The market research found a clear gap: no existing tool combined NLP for digitising and classifying complaints with LLM-generated reporting, in a package built around CQC standards. I designed an extended framework that adds LLM-based sentiment analysis and automatic report generation to fill that gap.

Outcome

The project delivered:

• the feasibility study and market research
• a working proof of concept with CQC-standard forms, a database, NLP classification, and a live dashboard
• an assessment of whether the system could become an off-the-shelf SaaS product for other care providers

The work also set out a proposed follow-on project, Complaint Prediction for Prevention (CPP-AI). It would use AI/ML to spot potential complaints before they happen, shifting from reacting to complaints to preventing them.

Tools

.NET Core, MySQL, Python (NLTK, spaCy), Node.js, PHP, Chart.js, REST APIs.