Inspiration

Finding a defect is often easier than finding out why it happened.

An operator notices damaged insulation on a cable. Maintenance knows which tool was used. The supplier knows the material revision. Quality has the applicable procedure. But that information is scattered across conversations, emails, documents, and tickets.

We wanted to connect these pieces without forcing operators to learn another application.

What if a WhatsApp report could start an investigatio, not just create another ticket?

That question inspired DéfecTive, a WhatsApp-first assistant that helps industrial teams collect evidence, investigate non-conformities, and coordinate a response. We chose an aerospace wiring-harness scenario for our MVP because it makes the distinction between seeing a defect and understanding its cause especially clear.

What it does

DéfecTive turns a photograph and a description sent through WhatsApp into a shared investigation.

The agent identifies visible observations, searches relevant technical documents and tickets, and determines what information is missing. It proposes targeted questions to maintenance through RCS and to suppliers through email, keeping their replies attached to the same dossier.

As answers arrive, the agent revises its hypotheses and prepares an investigation summary with supporting references, unresolved questions, and proposed next steps.

For example, an insulation nick might initially suggest several explanations. A maintenance reply identifying the installed tool, combined with a supplier response confirming the material family, can support a tooling-compatibility hypothesis. The photograph documents the symptom, the additional evidence helps explain it.

A quality reviewer inspects the evidence and approves the proposed action before operational guidance is returned through WhatsApp. Closure requires a separate effectiveness check, not simply an AI-generated answer.

Our interactive prototype demonstrates this complete journey using synthetic aerospace process records and participant responses. It does not provide certified aircraft maintenance instructions or authorize release to service.

How we built it

We designed DéfecTive as an event-driven AWS application, with an investigation agent hosted in Amazon Bedrock AgentCore.

AWS End User Messaging Social connects the WhatsApp channel to the backend. Amazon SNS distributes incoming events, Amazon SQS buffers processing, and AWS Lambda normalizes messages, associates them with dossiers, and applies the workflow rules. Amazon DynamoDB stores case state, evidence references, questions, approvals, and outgoing actions.

The agent uses Amazon Bedrock for image interpretation and structured investigation. It retrieves documentary evidence through Amazon Bedrock Knowledge Bases, backed by Amazon S3 and S3 Vectors, while current tickets are read separately from DynamoDB. This keeps documentary similarity distinct from authoritative operational status.

For collaboration, the implementation connects maintenance questions through AWS End User Messaging RCS and supplier correspondence through Amazon SES. Supplier replies are received into S3, parsed, and correlated with the relevant question.

The agent proposes questions and recommendations, but it does not directly control message delivery or approve its own conclusions. The backend checks recipients, permissions, case versions, and outgoing content before dispatch.

We developed the backend in Python with AWS SAM and CloudFormation templates for deployment. The quality portal uses Amazon Cognito, API Gateway, S3, and CloudFront.

The deliverables combine an interactive showcase, application source code, infrastructure templates, and local tests. Live provider integration and deployment verification remain separate from the simulated showcase.

Challenges we ran into

Keeping conversations connected. WhatsApp, RCS, and email do not share a common message format or conversation identifier. We needed explicit case and question identifiers, sender checks, and duplicate handling so that a reply would update the correct investigation.

Handling investigations that pause and resume. A supplier may answer long after the operator's report. Rather than keeping an agent running while waiting, we persist the dossier and resume processing when new evidence arrives. Late replies and uncertain delivery outcomes need their own handling—not just another retry.

Separating observations from conclusions. An image may show damaged insulation, but it cannot establish the tool used, the material history, or the underlying cause. We kept visual observations, attributed statements, documentary evidence, and hypotheses separate.

Making approval meaningful. Approval cannot remain valid indefinitely when the evidence changes. We tied it to a specific proposal and evidence revision, so contradictory information can invalidate an outdated decision.

Keeping the MVP focused. We prioritized one complete cable-defect journey over many disconnected features. The goal was to demonstrate useful investigation, not the largest possible collection of agents or services.

Accomplishments that we're proud of

We built an interactive prototype that brings together the operator, maintenance colleague, supplier, and quality reviewer in one understandable workflow.

The most important moment is not the first AI response. It is watching the investigation change after new evidence arrives, and seeing the application prevent an outdated approval from being reused.

Above all, we maintained a clear boundary: AI assists the investigation, qualified people retain decision authority.

What we learned

The most useful next step is often a question, not an answer.

An agent becomes more valuable when it can identify a missing fact, ask the appropriate person, and reconcile the response with existing evidence. Producing a longer explanation does not compensate for information that was never collected.

We also learned that messaging reliability is part of AI reliability. A well-written analysis is not useful if a response is attached to the wrong dossier, a message is duplicated, or an old instruction remains active.

Finally, human review works best when it is built into the workflow. Reviewers need accessible sources, explicit uncertainty, and a clear record of what changed—not just an approval button beneath a confident recommendation.

What's next for DéfecTive

Our next milestone is to validate the live communication channels and AWS deployment with authorized participants and rights-cleared images, then evaluate the workflow against manual coordination and document-only assistance.

We will measure investigation quality, citation correctness, time spent collecting information, and reviewer effort before making claims about operational savings.

Beyond the MVP, we plan to connect DéfecTive to quality-management systems, maintenance platforms, and ERP applications; expand multilingual support; and generate structured non-conformity and corrective-action reports.

The longer-term goal is to help teams recognize recurring problems and prevent the, not only respond after a defect appears.

One defect. One conversation. One coordinated path to resolution.