ReliefLens

Inspiration

During disasters, emergency coordinators receive photos, videos, and voice reports from multiple sources. These reports are often incomplete, contradictory, or difficult to verify in real time. Most AI systems summarize information, but they rarely communicate uncertainty or explain why a recommendation should or should not be trusted.

We wanted to build an AI assistant that behaves more like an experienced emergency coordinator—one that analyzes multimodal evidence, identifies conflicting information, asks the single most valuable clarification question, and only makes recommendations after critical uncertainty has been resolved.

That idea became ReliefLens.

What it does

ReliefLens is a multimodal AI-powered emergency reasoning assistant designed to support disaster response teams.

Users can upload:

📷 Photos
🎥 Videos
🎙️ Voice reports

ReliefLens analyzes every piece of evidence and produces:

Structured evidence extraction
Claim identification
Evidence strength scoring
Conflict detection
AI-generated clarification questions
Transparent reasoning summaries
Human verification workflow
Final recommendations backed by evidence

Unlike traditional AI assistants, ReliefLens intentionally withholds recommendations when critical uncertainty exists instead of making potentially unsafe assumptions.

How we built it

ReliefLens combines modern web technologies with multimodal AI.

Frontend

React
TypeScript
Tailwind CSS

Backend

Node.js
Express

AI

Google Gemini Multimodal API

Processing Pipeline

Photo
Video
Audio
      │
      ▼
Preprocessing
      │
      ▼
Gemini Analysis
      │
      ▼
Structured JSON
      │
      ▼
Schema Validation
      │
      ▼
Reasoning Engine
      │
      ▼
Conflict Detection
      │
      ▼
Clarification Question
      │
      ▼
Human Verification
      │
      ▼
Recommendation

To improve reliability, we implemented deterministic schema validation and fallback logic so the application continues functioning even when AI output is incomplete or uncertain.

Challenges we ran into

One of the biggest challenges was designing an AI system that doesn't simply generate an answer but reasons about uncertainty.

Some of the challenges included:

Handling multimodal inputs with different formats and sizes
Producing consistent structured outputs from AI responses
Detecting conflicting evidence across photos, videos, and audio
Preventing hallucinated recommendations
Building a transparent reasoning workflow that users can trust
Designing a deterministic fallback pipeline when AI confidence is low

Balancing flexibility with reliability required multiple iterations of both the prompting strategy and the validation pipeline.

Accomplishments that we're proud of

We're especially proud that ReliefLens doesn't just summarize evidence—it reasons about it.

Some highlights include:

Built a complete multimodal reasoning workflow
Developed transparent AI reasoning summaries
Implemented evidence conflict detection
Created an uncertainty-aware recommendation engine
Added human verification before final recommendations
Built deterministic fallback logic for improved robustness
Delivered a clean, intuitive interface suitable for emergency response scenarios

Most importantly, we built an AI system that knows when it needs more information before making important decisions.

What we learned

Building ReliefLens reinforced that trustworthy AI is not about generating answers as quickly as possible.

We learned:

Explainability is just as important as accuracy.
Human-in-the-loop workflows significantly improve trust.
Structured outputs make AI systems more reliable.
Validation layers are essential when working with generative AI.
Multimodal reasoning becomes much more valuable when evidence is transparent and traceable.

Perhaps our biggest takeaway was that uncertainty itself is useful information, and AI should communicate it rather than hide it.