At the SEMH provision I run, I been working on creating practical tools to help staff capture daily observations of children’s social, emotional, and mental health (SEMH) progress in a structured way. The aim is to ensure that day-to-day insights can feed into AI systems to generate weekly reports, 24-week (end of placement) progress reviews and support plans. 

Step 1 – Building the local LLM

Using Ollama and AnythingLLM, I built a local LLM and populated it with ‘dummy’ pupil data (exclusion reports, behaviour incidents reports and pupil profiles) to act as a proof of concept. This allowed me to analyse the pupil’s needs. I also ‘trained’ the data on the two frameworks (6 core strengths and Trauma Perceptive Practice (TPP)) we use to understand and support the additional SEMH needs of the children in our care, which provided further insights.

This means that the AI is only considering the data I add to the local LLM. In this case the ‘dummy’ pupil data, 6 core strengths and Trauma Perceptive Practice (TPP) frameworks.
 

Step 2: Capturing data for the local LLM

The next step was to think about practical, and effective ways of capturing data relating to the child. I wanted a workflow that allowed me to capture this information daily with minimum effort. I decided on using voice notes that could be transcribed by Otter AI into a text file that could then be added to the local LLM to ‘train’ it.

I needed a set of questions that I could answer to prompt me to give the required information allowing effective analysis of the child’s needs through the lens of the 6 core strengths and Trauma Perceptive Practice (TPP).

Using ChatGPT, I started by condensing the 6 Core Strengths framework into six daily questions each supported by guiding prompts. This would allow staff to reflect quickly without being overwhelmed. 
 
Also using AI, I then created a Trauma Perceptive Practice (TPP) checklist aligned with Compassion, Connection, and Belonging. I kept this short, simple, and consistent for daily use. 
 
To avoid duplication, I then combined both frameworks into a Unified Daily Checklist bringing the principles together into six broad areas covering attachment, self-regulation, participation, awareness, tolerance, and respect.

Finally, recognising the importance of documenting challenges as well as strengths, I added a seventh question for Challenging Behaviour allowing space to record incidents of verbal or physical abuse, disruption, or refusal. 
 
The result is a clean, one-page list of seven questions that can be used daily. Over time, these structured observations will feed into AI analysis (i.e the local LLM), making it easier to track patterns, highlight needs, and produce evidence for progress assessments. 
 
This is an exciting step toward harnessing AI to better support pupils’ SEMH development in a compassionate, consistent, and data-informed way. 

This is a work in progress. I am developing this approach as I learn.

Step 3: Considering what is next when building a local LLM

My next considerations are about data privacy, bias, and checking for accuracy before it is used on real pupil data.

If you have any questions, or are working on anything similar, please leave a comment below or contact me on the details below.