Delivering an AI-Powered Safeguarding Platform in 12 Weeks

Complete safeguarding system
including anonymous reporting, AI-powered risk assessment, and real-time analytics dashboards and messaging capabilities.

Only 12 weeks from concept to live production platform with the help of Anthropic’s Claude.

Faster delivery compared to traditional development methods. With AI handling code generation, this allowed engineers to focus on architecture and quality.

The background

Safeshout is a company that works with schools, youth services, and local authorities to close the gap between young people’s lived experiences and the systems designed to protect them. Safeshout created a safeguarding intelligence system with a powerful mission: giving young people a safe, trusted way to share concerns that might otherwise go unheard. Whether students choose to report anonymously or put their name to a concern, Safeshout turns their voice into real-time safeguarding intelligence, equipping schools and multi-academy trusts with the oversight, insight, and clarity to act with confidence.

The challenge

Safeshout partnered with Burendo to build a working prototype that could demonstrate their concept to schools. They needed a safeguarding and analytics platform built from the ground up, but as a startup, their limited funding meant traditional development approaches would be out of reach within their desired timeframe.

For Burendo, this represented an opportunity to showcase that innovative AI-assisted development could deliver a production-ready platform at pace whilst maintaining the standards of quality our clients expect.

Our approach

Burendo recognised that Safeshout’s vision could be achieved by leveraging AI-assisted development throughout the entire development lifecycle. We committed to using Claude (Anthropic’s AI assistant) strategically across requirement gathering, user story creation, prototyping, development, and ongoing iteration, not as a replacement for engineering expertise, but as a powerful augmentation tool that would allow our experienced team to work more efficiently.

To maintain robust governance and transparency, we established direct integration between Claude and Jira, enabling real-time status reporting and allowing user stories to be created and tracked directly within the project management system. This ensured that AI-assisted work remained visible, accountable, and aligned with traditional project governance frameworks.

Burendo integrated AI throughout the entire development lifecycle as a strategic engineering partner; something that could do the heavy lifting, but under the correct guidance and with full traceability through established project management tools.

Rapid prototyping and iteration
Claude generated initial infrastructure as code (AWS CDK) based on architectural requirements, creating the foundation for our serverless architecture in days rather than weeks. It built boilerplate React components and routing structure, allowing engineers to focus on complex business logic, and generated database schemas and API contracts from written specifications. What would typically take 3-4 weeks of setup was completed in under one week.

As the project progressed, the Burendo team maintained control of all critical decisions whilst leveraging Claude for execution speed. Senior engineers defined the architecture, security requirements, and technical approach. Product experts created detailed tickets with clear acceptance criteria and business context. Claude engaged in back-and-forth to understand the requirements fully and then generated initial code implementations based on these specifications. Engineers reviewed and refined the AI-generated code, ensuring it met quality standards and integrated properly. Claude assisted in debugging, optimisation, and handling edge cases.

Complex feature development
Claude assisted in building the dual-scoring AI triage system, implementing the integration between our custom machine-learning model and AWS Comprehend for sentiment analysis. It generated WebSocket connection handlers and state management logic for live dashboards and notifications, helped implement role-based access control across organisations with proper data isolation, and built data aggregation logic, heat map calculations, and exportable analytics report. Complex features that would typically take months were delivered in weeks, with engineers reviewing and refining AI-generated code rather than writing from scratch.

Continuous testing and quality assurance
To protect against regression and add confidence to the code that was being generated, we used Claude to also generate comprehensive test suites alongside feature development: unit tests, integration tests, and end-to-end test scenarios. Often working in a test-driven-development fashion, it created test data and edge case scenarios that engineers might not have considered and even embedded these checks and balances into automated path-to-live workflows, resulting in 28 CI/CD quality gates and release processes. Test coverage that would typically lag behind development was built in parallel, catching issues earlier and unlocking the confidence to enable continuous releases.

Documentation
Technical documentation was generated by Claude as features were built, creating API documentation, infrastructure diagrams, and onboarding guides, eliminating the bottlenecks that can occur in fast-moving projects. Doing this in sync with development meant that at any one time, the documentation was true, accurate, and comprehensive, which was critical as engineers rotated on or off the team, and as the point of handover approached. It also meant Claude itself remained up to date with critical decisions and features being implemented in previous sessions.

I have never onboarded to an engineering squad at this speed. Within a couple of days, I was making productive and valuable code contributions.
Kevin Kuszyk, Principal Consultant and Cloud Architect at Burendo

Technical architecture

Our team built a comprehensive AWS-native solution leveraging over 25 services across a multi-tenanted architecture designed specifically for the sensitive nature of safeguarding work:

  • Core infrastructure: Serverless architecture using Lambda, DynamoDB, S3, API Gateway, and WebSockets for real-time communication. We chose serverless to enable Safeshout to scale efficiently from their first school to nationwide deployment without the complexity and cost of managing physical servers.
  • AI/ML layer: Custom severity classifier integrated with AWS Comprehend for sentiment analysis and emotion detection. This included:
    • A dual scoring risk assessment process consisting of 2 parts. A scratch-built neural network custom severity classifier model built using PyTorch, NumPy & Pandas with over 16.8M parameters.
    • Trained using AWS SageMaker on example and human-labelled data provided by seasoned safeguard professionals, then deployed and continuously trained on live human sessions.
    • In addition, a custom algorithm scoring mechanism leveraging AWS Comprehend, textual, location, category, emotion and keyword analysis to determine a risk assessment.

By automating the initial assessment of concerns, we ensure that high-risk reports reach safeguarding teams immediately, potentially preventing harm to vulnerable students.

  • Security & authentication: Multi-layered authentication and encryption protecting separate pathways for staff and students. Given the sensitive nature of student disclosures, we implemented robust security measures to meet the stringent data protection requirements of educational safeguarding.
  • Frontend: React with TypeScript and Tailwind CSS, creating an intuitive interface hosted on AWS Amplify. We selected modern, proven technologies to create an interface that’s both reassuring for students in crisis and quick to load, removing barriers that might prevent a young person from reporting a concern.
  • DevOps: Infrastructure as Code with automated CI/CD pipelines and 28 quality assurance workflows throughout the path to live. Automated deployment pipelines allowed us to maintain rapid iteration and enabled unlimited ephemeral development environments and 3 path-to-live environments, whilst ensuring every release met security and performance standards before reaching production.

We conducted two AWS Well-Architected Reviews just three weeks apart to ensure best practices were met whilst considering security, performance, and maintainability. Burendo used AI to support these reviews with our Cloud Architect, creating a setup that would allow the Safeshout team to perform comprehensive assessments without requiring deep prior experience in formal architecture reviews. Between the two reviews, we used Claude to rapidly iterate on the findings, implementing improvements at a pace that would be impossible with traditional development and systematically resolving 5 of the 6 priority recommendations from the first review, resulting in a significantly more mature and production-ready architecture and raising the review score from 4 to 4.5.

Traditional agile processes are designed around typical development timelines. With AI dramatically accelerating code generation, we found that standard ceremonies like lengthy sprint planning sessions, refinements, and retrospectives weren’t suitable for the development pace. Instead, we maintained normal daily stand-ups and twice-weekly client demos, an unusually high cadence that kept stakeholders closely engaged with progress. This rapid feedback loop enabled Safeshout to see working features, provide immediate input, and watch their ideas come to life within days. Issues were caught early, priorities could shift quickly, and the client remained confident and excited throughout. The speed of AI-assisted development made this level of engagement possible, and the frequent touchpoints ensured we were always building the right thing.

How the platform protects students:
At the heart of our technical decisions was a deep understanding of the platform’s purpose: giving vulnerable young people a voice. We built three distinct reporting pathways to meet students where they are:

  • Logged-in accounts for students who feel comfortable being identified and want to track multiple concerns over time
  • Pseudonymous accounts using randomly generated usernames, allowing students to maintain continuity across multiple reports whilst protecting their identity from peers
  • Fully anonymous submission for students who need to report what they’ve witnessed or aren’t ready to be known yet

Each pathway flows into Safeshout’s intelligent triage system, where our custom AI models work alongside AWS Comprehend to assess severity, detect emotional distress, and surface urgent concerns automatically. This means safeguarding leads see high-risk reports immediately rather than them being lost in a queue, enabling schools to support vulnerable students as soon as they need it.

The outcomes

Platform capabilities delivered:

Super admin capabilities: for Safeshout’s team to onboard new organisations and fine-tune the AI algorithms as they learn from real-world usage.

Multi-tenanted role-based access control architecture supporting multiple organisations and multiple staff roles, each managing multiple schools with configurable locations, term dates, and settings.

Three reporting pathways designed with student psychology and safety in mind.

Intelligent concern assessment with automated AI-powered triage using a dual scoring system: sentiment analysis combined with a custom machine learning classifier trained on Safeshout’s safeguarding expertise.

Real-time dashboards with live feeds, comprehensive case management, messaging capabilities, and complete audit trails showing exactly who has accessed sensitive information, using WebSocket to provide immediate notifications and updates.

Rich analytics suite: heat-maps showing concern patterns across locations and times, concern category breakdowns, staff performance metrics tracking response times and workload, and system analytics monitoring processing performance, all exportable at the push of a button.

Evidence handling: secure file uploads allowing students to provide supporting evidence whilst maintaining anonymity.

Safeshout was consistently impressed with Burendo’s pace of delivery. Our twice-weekly demos regularly showcased 3 to 4 new features, with the rapid feedback loops allowing Safeshout to refine their vision in real-time and get the most from a fast-paced 12-week delivery timescale.

This work demonstrated that AI-assisted development, when guided by experienced engineers, can dramatically accelerate delivery whilst maintaining quality. Our methodology combines senior engineering expertise in requirements analysis, system architecture, and code quality with AI augmentation for code generation, testing, and documentation creation.

The key to success was maintaining human oversight at every critical decision point. Our engineers led solution design and architectural choices, using Claude as a force multiplier for implementation speed. AI is not a replacement for good engineers, it’s a tool to aid their development, whilst they provide the direction.

This was an ambitious undertaking. We committed to significantly more scope than traditional timelines would allow, backed by a target-based statement of work that protected the client whilst Burendo absorbed any risk. Having confidence in our AI-assisted approach, we not only met our commitments but exceeded them, delivering more capabilities than initially requested and progressing the application beyond a prototype into a production-ready platform.

With the right approach, combining experienced engineering oversight with AI acceleration, ambitious timelines become achievable without compromising on quality.

You’re making my dreams come true in front of my eyes!
Maddy Holder, Founder of Safeshout