SuperShield

(1) Inspiration: Buying a franchise can consume years of a family’s savings. Yet prospective owners must evaluate promotional claims, disclosure documents, financial projections, lease terms, and location risks—often without the analyst teams available to large companies. We built SuperShield because people making irreversible decisions deserve more than another chatbot. They need an agent that investigates independently, challenges unsupported claims, quantifies risk, and knows when to stop and ask a human.

(2) What it does: SuperShield is a human-controlled decision firewall for prospective franchise buyers. It:

Extracts claims and evidence from disclosure documents.
Detects contradictions and missing information.
Runs deterministic break-even, cash-runway, and downside calculations.
Tests findings against the buyer’s budget and personal constraints.
Rejects unsupported conclusions.
Drafts precise follow-up questions.
Requires human approval before any external action.
Produces an auditable Decision Packet with citations, calculations, uncertainties, and next steps. SuperShield performs the investigation; the human retains the decision.

(3) How we built it: SuperShield uses the Strands Agents SDK to coordinate a supervisor and specialized capabilities for evidence collection, adversarial review, financial analysis, validation, and human approval. The agent runs on Amazon Bedrock AgentCore Runtime and uses Amazon Bedrock for reasoning. Deterministic Python tools perform all financial calculations. A React interface displays the investigation plan, agent activity, citations, risks, approval checkpoints, and final Decision Packet. Amazon CloudWatch and AgentCore tracing provide observability. The public demonstration uses fictional documents and synthetic data to protect privacy and ensure reproducible results.

(4) Challenges we ran into: The hardest challenge was ensuring SuperShield behaved like a responsible agent rather than a document chatbot. We had to:

Separate retrieved document content from trusted instructions.
Prevent prompt injection embedded inside uploaded material.
Require evidence for every material conclusion.
Keep financial arithmetic outside the language model.
Make specialist agents genuinely distinct rather than role-play personas.
Preserve human authority over consequential actions.
Balance technical depth with a simple, understandable product experience.
Build a stable, inexpensive public demonstration within the hackathon deadline.

(5) Accomplishments that we're proud of: We are proud that SuperShield can complete a full investigation while remaining transparent about uncertainty. It can:

Connect claims to exact source passages.
Detect contradictions across multiple documents.
Quantify the financial consequences of overlooked terms.
Refuse to clear a case when required evidence is missing.
Ignore malicious instructions embedded in source material.
Recalculate only affected findings when assumptions change.
Pause at a policy-enforced human approval checkpoint.
Expose its Strands tool activity and AgentCore execution trace. In our evaluation, SuperShield achieved [insert verified results] across citation accuracy, seeded-risk detection, calculation correctness, and approval enforcement.

(6) What we learned: We learned that trustworthy agency is not created by giving several prompts different job titles. A useful agent needs bounded tools, durable state, evidence requirements, deterministic calculations, explicit permissions, failure handling, and the ability to abstain. We also learned that human control is not a limitation on autonomy—it is what makes autonomy appropriate for high-stakes decisions.

(8) What’s next for SuperShield: Next, we plan to:

Support additional business and investment document types.
Integrate licensed market and location datasets.
Add secure collaboration with accountants and legal advisers.
Monitor deadlines, document amendments, and unresolved evidence.
Expand the evaluation suite with expert-reviewed cases.
Introduce privacy-preserving multilingual explanations.
Develop production-grade identity, retention, and compliance controls. Our long-term goal is to make institutional-quality decision support accessible to every entrepreneur—not to replace their judgment, but to protect it.