AIdeas: RecoverMate
RecoverMate is an AI-powered "Recovery Concierge" that helps Australian disaster victims cut through the "second disaster" of overwhelming administrative bureaucracy. By leveraging AWS Generative AI to instantly decode complex insurance policies and match users to eligible government grants, we turn months of paperwork into a streamlined, actionable path to financial recovery.
App Category
Social Impact
My Vision
You survive the flames. You survive the flood. But then comes the real devastation, a 100 page insurance policy dropped in your lap, government websites that don't talk to each other, and weeks proving you lost what everyone can see is gone.
Poorly managed recovery is often identified as ‘the second disaster’; the unnecessary one. Anecdotal evidence indicates that the quality of a person’s recovery depends more on how well they manage the recovery years than what the event does. By Dr. Rob Gordon
My vision is simple, eliminate the bureaucracy of survival entirely.
RecoverMate isn't another AI chatbot spitting advice. I built a Recovery Concierge that takes action as a protective shield between shattered families and weaponized paperwork.
Here's what victims experience:
- Upload chaos (damage photos, muddy receipts, insurance PDFs)
- Instant clarity: "Your policy covers flash flooding. Here's exactly what you're owed." **
- Zero decisions: Grants auto-matched, forms pre-filled, documents organized
- One tap: "Review & Submit" replaces weeks of hell
A farmer sees: "Your income loss qualifies for $15K state relief + 13 weeks federal support." **
A family reads: "Good news your home qualifies under clause 4.2. Claim ready." **
A business owner gets: "3 grants matched. All forms complete. Submit now?" **
For agencies: Clean, perfect claims arrive, no back-and-forth, no errors, no panic calls.
For victims: Dignity returned. Time reclaimed. Mental space restored.
A family reads: "Good news your home qualifies under clause 4.2. Claim ready." **
A business owner gets: "3 grants matched. All forms complete. Submit now?" **
For agencies: Clean, perfect claims arrive, no back-and-forth, no errors, no panic calls.
For victims: Dignity returned. Time reclaimed. Mental space restored.
RecoverMate flips the system, AI absorbs every ounce of complexity so no Australian ever fights paperwork while standing in their ruins again.
Communities don't lack resilience, they lack bandwidth. This is cognitive infrastructure for national survival.
Why this matters
In emergencies, recovery implies returning a person, family or community to a previous state of being, covered or immersed in the protective medium of a familiar world. By Dr. Rob Gordon
RecoverMate fits perfectly in the Social Impact category because it doesn't just help individuals it rebuilds entire communities when Australia's most vulnerable need it most. Disaster recovery shouldn't be a second punishment.
- Millions in aid goes unclaimed: Every year, eligible families, farmers, and small businesses forfeit disaster funding because 100 page insurance policies and complex government forms are simply too overwhelming in crisis. RecoverMate instantly decodes Product Disclosure Statements, matches eligibility rules, and surfaces "hidden money" victims are legally entitled to but never find.
- Trauma needs cognitive relief: Research shows disaster shock destroys mental bandwidth, yet recovery demands peak cognitive function to navigate acronyms, portals, and exclusion clauses. RecoverMate becomes the "brain" victims lack shifting dense legal interpretation onto AI so families can focus on healing, not deciphering "storm surge vs riverine flood."
- Communities recover faster: When claims process in under 10 minutes instead of weeks, cash flows immediately tradespeople get hired to clear debris, rental properties get repaired, local shops reopen instead of shuttering permanently. This breaks post-disaster poverty cycles and injects liquidity when communities need it most.
- Agencies serve better: Government call centers get overwhelmed with panic queries after disasters. RecoverMate delivers clean, structured, pre-validated claims exactly matching Services Australia requirements targeting 30% call deflection so social workers focus on severe cases, not explaining basic eligibility 1,000 times.
- Cascading disasters demand new infrastructure: Australia faces back-to-back crises thus "Disaster fatigue" is real. RecoverMate proves Generative AI belongs in national resilience not as a corporate productivity gimmick, but as lifeline infrastructure that scales compassionately when human systems crack.
RecoverMate matters because it acts as a cognitive shield. It bridges the massive gap between private insurance constraints and public grant eligibility, instantly finding the "hidden money" victims are legally entitled to. It proves that we can use AWS not just for raw compute power, but for profound cognitive offloading/shifting the focus from fighting paperwork to rebuilding communities.
How I build this
Technical Approach: Spec-Driven Development with Kiro
I utilized Kiro as my primary development environment to bridge the gap between high-level policy analysis and low-level code execution. By leveraging Kiro’s spec-driven workflow, I was able to rapidly prototype the "Legal Translation Layer." I began by creating steering documents that mapped Australian insurance definitions (like "Flash Flood") to specific extraction prompts. This ensured that the AI didn't just "guess" but followed a strict logical framework for eligibility.
The Architecture: Serverless Meets Relational
The system is designed to be "Elastic" scaling to zero when not in use, and scaling to thousands of concurrent victims during a national disaster.
| Service | Description |
|---|---|
| Amazon Textract | Instantly digitizes messy, mud-stained receipts and multi-page PDFs with high accuracy. |
| Amazon Bedrock | Uses Claude 3 (Anthropic) to decode complex PDS legal jargon into empathetic, plain English. |
| AWS Lambda | Programmatically maps extracted data to official claim forms using pdf-lib. |
| Amazon S3 | Securely stores sensitive victim documents with encryption-at-rest via AWS KMS. |
| Amazon Cognito | Ensures HIPAA-level privacy and secure authentication for traumatized users. |
| Amazon RDS | Secure database with relational mapping for RecoverMate |
Figure 1 - Deployment Architecture
Security & Privacy-First Design
In a disaster recovery context, a data leak is a "second disaster." I implemented Defense-in-Depth to ensure victim data is untouchable.
Row-Level Security (RLS) via RDS
I chose Amazon RDS (Postgres) specifically for its Row-Level Security capabilities.
How it works: Every table (Claims, Receipts, Users) has a mandatory user_id column. I configured Postgres policies that check the app.current_user_id session variable.
The Result: Even if a bug existed in the application code, the database itself would refuse to return data unless the session identity matches the row owner. This is a hard-coded privacy guarantee.
How it works: Every table (Claims, Receipts, Users) has a mandatory user_id column. I configured Postgres policies that check the app.current_user_id session variable.
The Result: Even if a bug existed in the application code, the database itself would refuse to return data unless the session identity matches the row owner. This is a hard-coded privacy guarantee.
S3 Encryption at Rest
All documents uploaded to the "Smart Vault" (S3) are protected by AES-256 Server-Side Encryption (SSE-KMS).
KMS Integration: AWS Key Management Service (KMS) to manage the encryption keys.
Access Control: Access to these objects is restricted via IAM Roles and S3 Bucket Policies, ensuring only the specific Lambda function processing the claim can "see" the raw files.
KMS Integration: AWS Key Management Service (KMS) to manage the encryption keys.
Access Control: Access to these objects is restricted via IAM Roles and S3 Bucket Policies, ensuring only the specific Lambda function processing the claim can "see" the raw files.
The "Brain" - thinking process of the RecoveryMate
1. The "Eager" Extraction Pipeline
I designed a decoupled, event-driven ingestion engine. Instead of processing documents on-demand, RecoverMate processes them the moment they hit the cloud:
The S3 Trigger: When a user uploads a file, Amazon S3 fires an event that triggers an AWS Lambda function.
The One-Time Read: This Lambda immediately calls Amazon Textract to extract all raw text, tables, and form data.
Persistence Layer: The raw extracted JSON from Textract is stored in Amazon RDS (metadata and structure) and a "Processed" folder in S3 (full text). This becomes our "Single Source of Truth."
The S3 Trigger: When a user uploads a file, Amazon S3 fires an event that triggers an AWS Lambda function.
The One-Time Read: This Lambda immediately calls Amazon Textract to extract all raw text, tables, and form data.
Persistence Layer: The raw extracted JSON from Textract is stored in Amazon RDS (metadata and structure) and a "Processed" folder in S3 (full text). This becomes our "Single Source of Truth."
2. The "Context Injection" Stage
When the user actually initiates a claim, app no longer need to look at the original image.
Hydrating the Brain: Application pull the already-extracted data from RDS.
Bedrock Reasoning: This structured text is injected into the Amazon Bedrock prompt. Because the text is already clean and digitized, the LLM (Claude 3) can perform its "Legal Translation" almost instantly, mapping the data to the JSON schema required for the claim.
Hydrating the Brain: Application pull the already-extracted data from RDS.
Bedrock Reasoning: This structured text is injected into the Amazon Bedrock prompt. Because the text is already clean and digitized, the LLM (Claude 3) can perform its "Legal Translation" almost instantly, mapping the data to the JSON schema required for the claim.
Demo
Image 1: The landing page of RecoverMate serves as the central entry point for all users. From here, users can click the “Get Started” button to begin the claim assistance process.
Image 2: The Document Vault allows users to upload essential documents such as Product Disclosure Statements (PDS), receipts, insurance details, and other PDF-based materials. These files are processed using the AWS Text Extraction Service, which reads and extracts textual content. The extracted data is then securely stored in an AWS RDS PostgreSQL database.
Image 3: The first step of the Claim Assistance process enables users to select previously uploaded documents from the Document Vault. This information is used for the context injection mechanism, ensuring that relevant document content informs subsequent decision-making.
Image 4: In the second step, users provide additional contextual details including their property address, disaster type, and incident date. These inputs are utilized alongside the uploaded documentation to assess eligibility for government grants and insurance claims.
Image 5: The third step presents users with a list of eligible grants, if available. Additionally, a summary of insurance coverage is displayed, helping users understand possible claim options based on the provided and extracted data.
Image 6: The final step allows users to generate a claim form. This form can either be downloaded as a PDF or submitted directly , completing the claim assistance workflow.
What I learned
Building RecoverMate wasn't just a coding challenge; it was a lesson in how to build responsible, production-ready AI for a high-stakes social impact environment. Here are my key takeaways:
Spec-Driven Development vs. "Vibe Coding"
Before using Kiro, I was used to "vibe coding" sending random prompts to an AI and hoping for the best. Kiro forced a fundamental shift in my methodology. By using Kiro’s Spec-Driven workflow, I learned to decouple planning from execution.
Instead of writing code immediately, I spent time refining Requirements and Technical Design documents. This structured approach meant that when the AI started building, it already understood the complex relationship between Australian insurance definitions and government eligibility criteria, drastically reducing bugs and "hallucinations."
Privacy as a Feature, Not a Chore
Handling trauma data (IDs, bank details, damage photos) is a massive responsibility. I learned that in a social impact project, security is your most important user feature.
Utilizing Amazon Bedrock Guardrails and PII Redaction, We will ensured that sensitive victim data will never be used for model training and remained within a secure, encrypted AWS environment. Building a system that users can trust during a crisis is just as important as building a system that works.
Serverless is the best Choice for Disaster Tech
Disaster recovery applications face a unique "Spiky Load" problem, they sit at zero traffic for months, then spike to thousands of concurrent users in a single day.
I learned how to optimize AWS Lambda On-Demand to handle these bursts without manual intervention. Serverless isn't just a cost-saver, it’s a resilience strategy.
Future Improvements
- Multimodal Evidence Processing: Expand beyond text to include image analysis. By using Amazon Bedrock (Claude 3 Vision), the system will analyze photos of property damage to automatically categorize losses and cross-reference them with insurance "inclusion" clauses.
- Direct Insurance API Integration: Transition from generating PDFs to direct data transmission. Following approval from major insurers, RecoverMate will use secure APIs to push claim data straight into their back-office systems, bypassing manual paperwork entirely.
- Government Grant Automation: Partner with agencies like Services Australia to utilize official APIs. This will allow the "Action Engine" to submit grant applications directly to government portals, ensuring victims receive emergency liquidity without navigating complex myGov forms.
References & Further Reading
- The "Second Disaster" Defined: Australian Journal of Emergency Management (AJEM). "Recovery: The Second Disaster." An analysis of how bureaucratic friction and poorly managed recovery cycles cause secondary trauma. View Publication
- The $2.6 Billion Unclaimed Gap: ASIC (Australian Securities and Investments Commission). "Unclaimed Money Search." Statistics on the billions in lost or unclaimed insurance and bank funds currently held by the Australian Government. View Data
- The Underinsurance Crisis: The Australia Institute. "Climate change, home values, and underinsurance." Research identifying that 800,000 Australian households are at risk of losing up to 74% of their total wealth due to insurance navigation barriers. Read Report
- Systemic Claims Issues: Financial Rights Legal Centre (FRLC). "Exposed: Claims Handling in the Disaster Recovery Sector." A report highlighting that complex policy jargon and poor claims handling are the #1 issue facing Australian disaster victims. Read Report
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article