As educators, we are constantly seeking ways to provide immediate, specific, and effective support for our pupils, especially those with Social, Emotional, and Mental Health (SEMH) needs. We know the dedication is there, but sometimes, the sheer volume of qualitative data—from incident reports to specialist advice—can feel overwhelming.

The question is: What if every child had a dedicated AI assistant helping schools understand and support their individual requirements?

I’ve been developing a prototype using a Local Large Language Model (LMM) to address this very challenge. This innovative approach moves beyond basic generative AI tasks to provide deep, secure SEMH data analysis that could revolutionise how we operate. Let’s dive into how this framework works and why it’s a game-changer for Local LLM SEMH support in schools.

---

The Data Challenge: Why Current Methods Fall Short

Despite the incredibly hard work of professionals, many schools are currently ineffective at collecting, analysing, and using SEND data. Valuable insights are often buried in paper files, scattered across different systems, or simply forgotten as staff change.

This lack of centralised, easily accessible data means:

• Support is reactive, not proactive.
• New staff struggle to quickly get up to speed on a child’s history.
• The compound impact of professional advice (like reports from Educational Psychologists) diminishes over time because the information isn’t consistently being applied or updated.

We need a solution that captures this organisational memory and makes it instantly actionable.

---

What Exactly is a Local LLM for Education?

When we talk about a Local LLM, we’re not discussing a public, cloud-based tool like ChatGPT. We are referring to a private and secure piece of software trained only on the specific data you feed it.

• Advanced Pattern Recognition: An LLM is excellent at recognising complex patterns in text. By training it on pupil data, it learns the unique behavioural triggers, support strategies, and historical context for each child.
• Security and Trust: Crucially, a local model does not send your sensitive pupil information to external servers. It only responds to and uses the data in your secure system, significantly reducing the risk of “hallucination” or providing irrelevant answers.

This targeted training means the AI’s output is entirely based on the pupil’s specific documents, leading to higher confidence in the advice provided.

---

The Game-Changing Benefits of an AI ‘Organisational Memory’ 🧠

Implementing a Local LLM provides a range of benefits that traditional data systems simply cannot match:

1. Massively Impacted Organisational Memory

The LMM won’t forget any data it has been given. Advice received from a specialist two years ago remains in the system, constantly being used and updated in all subsequent analysis.

2. Effective Staff Onboarding

When a new Teaching Assistant (TA) or teacher joins, they no longer need to sift through mountains of paperwork. They can simply ask the LMM: “What are the three most important things I need to know about Bruce to support him today?”—and get an immediate, data-backed answer.

3. 24/7 Access to Professional Guidance

No more waiting for a professional to be available. Staff have instant access to tailored advice, whether they are planning a lesson at 6 PM or preparing for a sports day adaptation at 8 AM.

Three examples of using a local LLM in schools

Scenario 1: New Teaching Assistant supporting pupil for the first time

Prompt:

“Using only the documents I have uploaded to AnythingLLM, I am a new member of staff supporting Bruce Banner, what are his top 3 triggers for violent outbursts and what strategy do you recommend to support him regulate please”

---

Scenario 2: Teacher planning sports day and needs to understand what adaptations need to be made to support the pupil

Prompt:

“Using the documents I have uploaded to AnythingLLM, please identify the adaption to support Bruce Banner that I will need to make to an activity where he will be competing against other children in sports day.”

Scenario 3: Senior Leader wants to prioritise the elements that staff are trained in Trauma Perceptive Practice (TPP) based on the additional SEMH needs of the pupil

Prompt:

“Based on the ‘ZZ Trauma Perceptive Practice Element Summary’ document I have uploaded to this chat only, and the documents I have uploaded to AnythingLLM, staff at my school are trained in different elements of Trauma Perceptive Practice. Which are the most important elements a staff must be trained and well versed in to support Bruce Banner and why? Please rank the elements in order that support Bruce Banner, from most to least. Do not use any information or knowledge outside of these parameters.”

A Practical Workflow for Sustainable Implementation

To make Local LLM SEMH support a reality, we need an administrative process that is both effective and low-effort. Here is a suggested workflow to execute this approach:

1. Daily Staff Insights: Staff provide daily qualitative observations and insights, ideally in line with existing frameworks (e.g., Six Core Strengths, TPP) [Internal Link to School’s SEND Policy]. To minimise admin, this should be a quick voice recording (e.g., using an app like Otter AI).
2. Auto-Transcription: The recording is automatically transcribed by an AI.
3. LMM Ingestion: The transcribed text is fed directly into the secure local LMM, alongside all existing pupil data and the specific SEND frameworks your school operates under.
4. Specialist & Carer Input: Reports from Educational Psychologists, Inclusion Partners, and feedback from parents and carers are also ingested, adding extra layers of detail and context for analysis.
5. Output & Action: Staff write simple prompts to get various outcomes:
   • “Based on Bruce’s last three incidents, what staff training should we prioritise this term?”
   • “What adaptions are necessary for Nick to successfully participate in the upcoming activity?”

This method ensures we are capturing insights with minimal administrative burden and using that rich data to generate concrete, actionable support plans for every pupil.

---

The Future of Targeted Support

The use of a Local LLM for data analysis is a paradigm shift in how we approach SEND support. It empowers school staff with an intelligent, ever-present tool that ensures no child’s needs are forgotten and every intervention is data-informed.

The prototype proves the concept is viable and accessible. The time has come to leverage this technology to build a truly comprehensive, constantly updating support system for our most vulnerable pupils.

---

About the Author

Enzo Vullo is a former SENCo and now Head of a SEMH provision with an interest in EdTech and AI . Currently he is exploring AI in education, with a ficus of building local LLMs and using prompt engineering to analyse qualitative pupil data to help pupils progress. You can learn more about Enzo’s work and connect with him on his LinkedIn profile.