Document Triage
Incoming correspondence scanned, text extracted, and risk classified as LOW, MEDIUM, HIGH, or CRITICAL within seconds.
AWS AIdeas 2025 · Top 50 Finalist · Social Impact
AI that advocates for the vulnerable — before it's too late
CivicGuardian AI is a multi-agent safeguarding platform built on Amazon Bedrock Nova that protects vulnerable adults from missing critical deadlines in housing, benefits, and healthcare correspondence. It reads what arrives, identifies the risk, and alerts the right person before a deadline is missed. Top 50 Finalist in the AWS Global 10,000 AIdeas Competition.
CivicGuardian · Live Safeguarding Alerts
Static demo preview
Demo case · Human review required
HIGH RISK · IMMEDIATE REVIEW REQUIRED
What CivicGuardian found:
Missed welfare calls, GP weight-loss flag, and possible self-neglect pattern.
What happens next:
Risk Analyst classifies the case as HIGH. Policy Reasoner prepares a draft response. Governor checks safety, tone, deadline logic, and escalation requirements.
Human review: Waiting for caseworker approval before any action.
ELEVATED · REVIEW WITHIN 48 HOURS
What CivicGuardian found:
Carer stress indicators across 3 contact points.
What happens next:
The system recommends a carer support assessment before the next visit.
Human review: Draft recommendation gated for human review.
MONITORING · NO IMMEDIATE ACTION
What CivicGuardian found:
Medication adherence improved after the last intervention.
What happens next:
Low-risk case archived with audit trail. Policy Reasoner skipped to control cost.
Human review: Available for caseworker review if needed.
Guardian Loop: risk classified, context checked, response gated for human approval. No case is silently dropped.
Why CivicGuardian exists
The information is there. The system just doesn't read it.
Margaret is 72. She has early-stage dementia. She receives 15 letters a month from her council, her housing association, and her NHS trust. Last month she received a housing benefit review notice with a 14-day deadline buried in the third paragraph. She missed it. Two weeks later she received an eviction warning.
Margaret is not an edge case. She is one of 12,000 vulnerable adults in the UK who lose housing every year due to missed correspondence deadlines. Her caseworker manages 40 or more clients. There is no system in place that reads Margaret's post, identifies the risk, and alerts her caseworker before the deadline passes.
Every existing system in UK social care technology is either internal or passive. Nourish manages what happens inside the care relationship. LiquidLogic and Mosaic manage statutory workflows inside local authority systems. Citizens Advice provides information when someone asks for it. None of these systems monitor correspondence that arrives from outside the care relationship. None of them act before a deadline is missed.
CivicGuardian AI is the only system that is both external-facing and proactively protective.
12,000+
Vulnerable adults lose housing annually due to missed deadlines
15+
Letters per month received by a typical vulnerable adult
40+
Clients managed per caseworker
0
Existing systems that read incoming correspondence proactively
How It Works
Intelligence that works continuously, so your team doesn't have to
Ingest
Correspondence arrives in the secure inbox. Text is extracted, consent status is checked, and the Guardian Loop begins.
Analyse
Three AI agents process the document in sequence. The Risk Analyst triages. The Policy Reasoner drafts a response. The Governor validates before anything reaches a human.
Alert
The caseworker receives a colour-coded notification with the risk level, deadline, AI-drafted response, and one-click approve and reject buttons.
Close
Every decision is recorded. Every case has a full audit trail. Nothing leaves the system without human approval. Nothing is ever silently dropped.
Capabilities
Built for the complexity of real social care
Risk Stratification
Dynamically scores risk across the caseload. Updated with every new document, not at review intervals.
Deadline Detection
Identifies critical deadlines buried in complex correspondence. Benefit reviews, hearing dates, eviction notices. Flagged before they pass.
Policy Response Drafting
The Policy Reasoner agent drafts legally-grounded responses citing UK legislation. Housing Benefit Regulations 2006, Care Act 2014, Mental Capacity Act 2005.
Audit Trail
Every recommendation logged with full evidence and reasoning. Supports serious case reviews and demonstrates due diligence.
Consent Management
Pipeline checks consent status before processing any document. Withdrawal halts processing immediately and triggers PII deletion. Mandatory AI disclosure on every output.
Sectors
Built for the organisations that carry the most responsibility
Local Authorities
Adult social care, high-volume caseloads. CivicGuardian restores early intervention capacity without increasing headcount.
NHS Trusts
Community health and mental health trusts. Cross-pathway risk signals integrated from NHS systems.
Housing Associations
Property-level signals (repair reports, access issues, rent arrears) integrated with social care data.
Technology
Enterprise-grade AI, built for public sector trust
CivicGuardian is built on Amazon Bedrock Nova — AWS's most capable foundation model for enterprise workloads. All data remains in the UK (eu-west-2, London). No cross-border transfers. No model training on your data. Amazon Bedrock's architecture guarantees your case data is never used to improve the underlying model.
The system is deployed in eu-west-2 (London). A deliberate decision, not a default. UK citizen data stays in the UK, aligned with UK GDPR, the Care Act 2014, and NHS Digital data residency requirements.
Amazon Bedrock Nova
AWS UK Region
ISO 27001 aligned
NHS DSP Toolkit
UK GDPR
REST API
Event-driven architecture
Zero-trust security
Compliance commitments
Data residency UK only
All data in eu-west-2 (London), no cross-border transfers.
No model training on your data
Amazon Bedrock guarantee — your data never improves the base model.
Human decision primacy
AI informs, practitioners decide. No automated actions.
Full explainability
Every alert shows its evidence and reasoning. No black boxes.
The Guardian Loop
Three agents. One safety-first cycle.
CivicGuardian doesn't replace human judgement — it extends it. A serverless multi-agent system processes every piece of correspondence through a continuous, safety-first loop before any human review.
Risk Analyst
Amazon Nova Lite
Powered by Amazon Bedrock Nova Lite. Triages every document and classifies risk as LOW, MEDIUM, HIGH, or CRITICAL.
Policy Reasoner
Amazon Nova Pro
Powered by Amazon Bedrock Nova Pro. Drafts a legally-grounded response citing UK legislation, but only for MEDIUM risk and above. This is enforced at the Step Functions CostGate Choice state. Every case record in DynamoDB contains a nova_pro_called Boolean field as evidence.
Governor
Nova Lite
Powered by Nova Lite. Validates every draft for factual consistency, legal soundness, tone, and deadline logic. If confidence falls below 0.75, escalation is mandatory regardless of risk level.
AWS Infrastructure
Orchestration
Amazon Bedrock + Step Functions
Compute
AWS Lambda (serverless)
Storage
S3 (AES-256) + DynamoDB
Escalation
Amazon SNS + CloudWatch
Security
Secrets Manager + IAM least-privilege
Validated Performance
Avg response time
<3 seconds
Classification accuracy
94.2%
Unit tests passing
122 / 122
Multi-agent accuracy gain
+92%
Cost reduction vs single-model
42%
Compliance & Governance
Built for the most sensitive data in public services
Vulnerable adults' case data is among the most sensitive information in existence. CivicGuardian was designed with a privacy-first architecture from day one — not retrofitted for compliance.
Every design decision, from encryption at rest and lifecycle-based retention to deterministic Governor validation, was made to ensure the system could be trusted by practitioners, commissioners, and data protection officers alike.
Privacy Framework
Both S3 buckets use AES-256 encryption with public access blocked. A 90-day lifecycle policy transitions documents to Glacier and expires them after seven years. CloudWatch log sanitisation ensures raw letter content never reaches logs. All data stays in eu-west-2.
Consent Framework
Aligned with the UK Mental Capacity Act 2005 and UK GDPR Article 9. The pipeline checks consent_status before processing any document. If withdrawn or pending, processing halts immediately. A dedicated consent withdrawal Lambda deletes all S3 documents, removes PII fields from DynamoDB, and notifies the caseworker.
Escalation Framework
SLA-enforced wait states. CRITICAL cases require caseworker response within 4 hours. HIGH within 24 hours. MEDIUM within 48 hours. If the caseworker does not respond, Step Functions automatically transitions to supervisor escalation. If the supervisor does not respond on a CRITICAL case, a final emergency alert fires. No case is ever silently dropped.
AI Disclosure
Every caseworker notification carries a mandatory label: "This response was generated by CivicGuardian AI and requires caseworker review and approval before any action is taken."
Roadmap
What CivicGuardian becomes next
Phase 2. Next phase, not yet built. The current system processes written correspondence. The next phase extends the Guardian Loop to audio, multilingual communication, and proactive outreach, removing every remaining barrier between a vulnerable adult and the help they need.
Audio Correspondence
Voicemail and audio message transcription via Amazon Transcribe. Spoken correspondence from service users processed through the same Guardian Loop as written documents — no channel left unmonitored.
Multilingual Translation
Real-time translation via Amazon Translate — processing correspondence in any language and generating responses in the service user's preferred language. Removing language as a barrier to safeguarding.
Plain-Language Summaries
Auto-generated plain-English summaries of complex legal letters — designed for service users with dementia, learning disabilities, or low literacy. Delivered alongside the original document.
Proactive Outreach
Automated deadline reminders via SMS and voice call — triggered when a high-risk deadline is approaching and no response has been logged. Turning reactive triage into proactive safeguarding.
EHR / Case System Integration
Direct API integration with Liquid Logic, Mosaic, and NHS systems — eliminating manual upload and enabling continuous monitoring of live case records without practitioner intervention.
Cross-Agency Risk Sharing
Secure, consent-based risk signal sharing between local authority, NHS, and housing association systems — enabling the multi-agency picture that currently only exists in serious case reviews.
Target Impact
The outcomes we are building toward
12,000+ vulnerable adults lose housing annually due to missed correspondence deadlines. Reactive crisis intervention costs 3–5x more than proactive safeguarding. CivicGuardian is designed to change both numbers.
<5%
Target missed deadline rate
80%
Caseworker approval target for AI drafts
3–5x
Cost saving vs reactive crisis intervention
12,000+
Vulnerable adults at risk annually in the UK
