AI-Powered Institutional Insights

Comprehensive guide to the AI-powered insights feature that transforms student wellbeing data into actionable institutional intelligence.

Overview

The AI Insights feature uses AWS Bedrock (Claude 3 Sonnet) to analyze aggregated student wellbeing data and generate comprehensive institutional reports. Instead of manually interpreting keywords and themes, university staff receive AI-generated analysis with specific patterns, affected cohorts, and actionable recommendations.

Deployment Date: December 2025
Status: Production
Location: Admin Dashboard → Overview → Student Concerns & Positive Themes
Cost: Approximately £12/month per university (1 daily analysis)

Key Features

Intelligent Theme Analysis

Identifies patterns across multiple data points
Groups related concerns into coherent themes
Calculates severity levels based on student count and risk indicators
Tracks trends over time (increasing/stable/decreasing)

Cohort Identification

Determines which student groups are most affected
Examples: "First years", "International students", "STEM students"
Helps target interventions to specific populations

Actionable Recommendations

Provides 3-4 specific, evidence-based actions for staff
Tailored to the actual concerns identified
Practical and implementable at institutional level

Professional Reporting

Generates institutional-quality analysis
Exportable as text reports for meetings
UK English, professional tone
Privacy-first (no individual student identification)

How to Use

Access the Feature

Navigate to admin.mindmeasure.co.uk (opens in a new tab)
Select your university from the dashboard
Go to the Overview tab
Scroll to "Student Concerns" or "Positive Themes" cards

Generate AI Insights

Click the purple "AI Insights" button on either card
Wait 10-15 seconds for analysis to complete
Review the comprehensive report in the modal
Export as text report if needed for meetings

Understanding the Output

Each AI-generated insight includes:

Theme Name

Professional identification of the concern or positive theme

Example: "Academic Pressure & Time Management"

Summary

2-3 sentence explanation of what students are experiencing

Written in clear, professional UK English
Contextualizes the raw data

Key Patterns

Specific patterns identified with percentages

Example: "67% mention multiple deadlines within same week"
Helps understand the nuances of the concern

Severity Assessment

Four-level rating system:

Critical: Immediate intervention required
High: Significant concern requiring prompt action
Medium: Notable issue to monitor and address
Low: Positive trend to maintain

Trend Direction

Increasing: Growing concern (↑ with percentage)
Stable: Consistent with previous periods (→)
Decreasing: Improving situation (↓ with percentage)

Affected Cohorts

Student groups most impacted by this theme

Helps target support services effectively
Examples: "First years during exam periods", "International students"

Recommended Actions

3-4 specific steps university staff can take

Evidence-based and practical
Aligned with the identified patterns
Implementable at institutional level

Related Resources

Types of support services that would help

Guides resource allocation
Identifies gaps in current offerings

Exporting Reports

Click "Export Report" button in the insights modal
Saves as plain text file with timestamp
Formatted for easy sharing in meetings
Contains all insights with full detail

Technical Architecture

Backend API Endpoint

Location: /api/ai-insights/generate
Method: POST
Authentication: Admin credentials via session

Request format:

{
  "type": "concern" | "positive",
  "data": [
    {
      "keyword": "overwhelmed",
      "count": 187,
      "severity": "high"
    }
  ],
  "universityId": "worcester",
  "timeRange": "30d"
}

Response format:

{
  "insights": [
    {
      "themeName": "Academic Pressure & Time Management",
      "summary": "Students are reporting...",
      "studentCount": 287,
      "severity": "high",
      "keyPatterns": [
        {
          "pattern": "Multiple deadlines same week",
          "percentage": 67
        }
      ],
      "trendDirection": "increasing",
      "trendPercentage": 23,
      "affectedCohorts": ["First years", "STEM students"],
      "recommendedActions": ["Review assessment scheduling..."],
      "relatedResources": ["Academic Skills Workshops"]
    }
  ],
  "generatedAt": "2025-12-15T10:30:00Z",
  "analysisScope": {
    "studentCount": 287,
    "timeFrame": "30d"
  }
}

AWS Bedrock Integration

Model: Claude 3 Sonnet (anthropic.claude-3-sonnet-20240229-v1:0)
Region: eu-west-2
Authentication: Server-side credentials (stored in Vercel environment variables)

Key configuration:

Max tokens: 4000
Temperature: 0.7 (balanced creativity and consistency)
System prompt: Specialized for institutional wellbeing analysis

Data Flow

Frontend → User clicks "AI Insights" button
Service Layer → AIInsightsService.ts prepares request
API Endpoint → /api/ai-insights/generate receives request
Prompt Construction → Builds specialized prompt with context
Bedrock Call → Invokes Claude 3 Sonnet via AWS SDK
Response Parsing → Validates and structures JSON response
Frontend Display → AIInsightsModal.tsx renders results

Security & Privacy

Data Protection

Aggregated data only (no individual transcripts sent to AI)
Keywords and counts, not personally identifiable information
Server-side processing (credentials never exposed to browser)
HTTPS encryption for all API calls

AWS Credentials

Stored as encrypted Vercel environment variables
Shared with mobile app's Bedrock integration
Scoped to bedrock:InvokeModel permission only
Automatically trimmed to prevent formatting errors

Compliance

GDPR compliant (no personal data processing)
HIPAA aligned (aggregated insights only)
Privacy-first design
No student identification in outputs

Cost Management

Pricing Structure

Claude 3 Sonnet Pricing:

Input: $3.00 per million tokens
Output: $15.00 per million tokens

Typical Analysis:

Input tokens: ~800 (keywords, context, system prompt)
Output tokens: ~1200 (insights JSON)
Cost per analysis: ~$0.02

Monthly Estimates:

Daily analysis: ~£12/month per university
On-demand only: ~£0.60/month (1-2 analyses per week)

Cost Optimization Options

Option 1: Claude 3 Haiku (60x cheaper)

Model: anthropic.claude-3-haiku-20240307-v1:0
Cost: ~$0.0003 per analysis
Monthly: ~£0.20 per university
Trade-off: Slightly less sophisticated analysis

Option 2: Scheduled Analysis

Generate insights once daily at 2am
Store in database for instant access
No waiting time for users
Predictable monthly costs

Option 3: Caching

Cache insights for 24 hours
Multiple users see same analysis
Reduced API calls
Cost savings for high-traffic dashboards

Troubleshooting

"The security token included in the request is invalid"

Cause: AWS credentials expired or incorrectly formatted
Solution:

Verify credentials in Vercel environment variables
Check for extra spaces or newlines
Ensure credentials match mobile app's working credentials
Redeploy after updating

"r.insights.map is not a function"

Cause: API response structure mismatch
Solution:

Check API returns { insights: [...] } not { insights: { insights: [...] } }
Verify parseClaudeResponse returns correct structure
Check browser console for actual response structure

Modal shows "No significant insights generated"

Cause: Claude returned empty or malformed JSON
Solution:

Check Vercel function logs for Claude's raw response
Verify prompt is constructed correctly
Ensure sufficient data provided (at least 3-5 keywords)
Check Claude didn't hit token limit

Analysis takes longer than 15 seconds

Cause: Bedrock throttling or network latency
Solution:

Check AWS Bedrock quotas for eu-west-2
Request quota increase if hitting limits
Consider implementing timeout handling
Add retry logic for transient failures

S3 uploads fail alongside AI Insights

Cause: Shared AWS credentials expired
Solution:

Update AWS credentials in Vercel (both projects share same credentials)
Test S3 uploads after fixing AI Insights
Both features use same IAM user

Best Practices

When to Generate Insights

Recommended:

Weekly review of student concerns
Monthly trend analysis
Before staff wellbeing meetings
When preparing institutional reports
After significant changes in patterns

Avoid:

Multiple times per day (data doesn't change that fast)
Without reviewing raw data first
As only source of information (complement with direct student feedback)

Interpreting Results

Do:

Cross-reference with raw keyword data
Consider institutional context
Share with relevant staff teams
Track recommendations implemented
Monitor impact of interventions

Don't:

Take recommendations as absolute directives
Ignore institutional constraints
Share publicly without review
Use for individual student identification
Rely solely on AI without professional judgment

Combining with Other Data

AI Insights work best when combined with:

Direct student feedback surveys
Staff observations and reports
Academic performance data
Support service utilization metrics
Engagement and attendance data

Future Enhancements

Planned Features

Phase B: Scheduled Analysis

Automatic daily generation at 2am
Store insights in database
Historical trend tracking
Email summaries to staff

Representative Quotes

Extract actual student quotes from transcripts
Add context to keyword patterns
Maintain privacy (anonymized)

Longitudinal Analysis

Compare insights across terms
Identify seasonal patterns
Track intervention effectiveness
Year-over-year comparisons

Custom Prompts

Institution-specific analysis focus
Configurable severity thresholds
Custom cohort definitions
Tailored recommendation types

API Reference

For detailed API specifications and code examples, see:

Backend API Documentation
Source code: mind-measure-core/api/ai-insights/generate.ts
Frontend service: mind-measure-core/src/services/AIInsightsService.ts
Modal component: mind-measure-core/src/components/AIInsightsModal.tsx

Questions or issues? Contact the development team or check the GitHub repository (opens in a new tab) for the latest updates.

Overview CMS User Guide