RELAY: What If AI Could Hand Over Knowledge Without Losing the Context? 🔗🤖

This is a submission for the MLH x DEV Writing Challenge

What I Built

Have you ever tried to hand over a task to someone else, only to realize that all the important details exist in your head, scattered messages, random notes, and old conversations?

Now imagine an AI system that could bring those details together, distinguish facts from uncertainty, and prepare a useful handoff without pretending to know things it doesn't.

That's the idea behind RELAY — Personal Knowledge Handoff AI, a project I built to explore a more reliable way of transferring knowledge.

RELAY takes messy personal knowledge, such as messages, notes, receipts, images, and voice notes, and organizes it into a temporary, evidence-backed handoff.

Instead of simply generating a summary, the project focuses on questions that matter when someone else needs to act on the information:

• Evidence-backed answers: Can we connect an answer to the information that supports it?
• Temporal conflict detection: What happens when two pieces of information disagree or refer to different points in time?
• Explicit uncertainty: Can the system acknowledge what it doesn't know instead of presenting guesses as facts?
• Source provenance: Where did a particular piece of information come from?
• Temporary handoffs: How can we transfer the context someone needs without treating every detail as permanent knowledge?

The goal is simple: make knowledge easier to transfer while preserving the context behind it.

Why I Built It

I wanted to explore a problem that goes beyond generating answers with AI.

In many real-world situations, the challenge isn't a lack of information. It's that the information is fragmented, outdated, contradictory, or difficult for another person to reconstruct.

A generic summary can miss these distinctions. A confident AI-generated answer can also be misleading when the underlying evidence is incomplete.

RELAY explores a different approach: the handoff should preserve evidence and uncertainty, not just produce a polished summary.

This project helped me think more deeply about how AI systems can be designed around trust, traceability, and practical usefulness rather than generation alone.

Demo

🎥 Watch the demo: https://youtu.be/5DhEMq-5jpc

💻 Explore the source code: https://github.com/Tanishaaaaaaa/relay-ai-handoff

The key idea behind RELAY is that the answer itself isn't the whole story. Being able to understand where information came from, identify conflicting details, and recognize uncertainty is equally important.

Partner Technologies

RELAY explores an open-weight, local-first AI architecture alongside an application designed to organize information and build evidence-backed handoffs.

1. Open-weight AI

The project explores using open-weight models for knowledge processing. This approach offers flexibility in how AI inference can be configured and deployed, including the possibility of running models locally.

2. Evidence and provenance

An important part of the project is connecting extracted information to its supporting sources. This makes the handoff more useful than a standalone generated summary and helps preserve the context behind individual claims.

3. Conflict detection and uncertainty

Information can change over time, and different sources can disagree. RELAY's design explicitly considers temporal conflicts and uncertainty instead of assuming every extracted fact is equally reliable.

4. Local-first architecture

The project explores a more flexible architecture in which local AI inference can be an option. This is especially interesting when considering the control developers may want over their AI infrastructure and data workflows.

The broader lesson for me was that choosing an AI model is only one part of building a useful AI application. The surrounding system — including evidence handling, conflict detection, and uncertainty — matters just as much.

For the MLH partner-technology requirement, I would identify the specific partner product and explain its actual implementation here before submitting. I don't want to claim an integration that the project hasn't verified.

Hackathon Experience

Building projects around AI is a great reminder that an interesting idea is only the beginning.

The real work comes from turning that idea into a concrete problem, making architectural decisions, experimenting with tools, and deciding which details deserve the most attention.

While working on RELAY, I kept coming back to one question:

What makes an AI system useful enough to trust with a real task?

For me, the answer isn't just a better model. It's a system that can explain its answers, preserve the source of its information, recognize contradictions, and be honest about what remains unknown.

That perspective shaped how I approached the project and what I wanted to learn from building it.

I'm excited to keep exploring AI systems that solve practical problems while taking reliability and context seriously.

Thanks to MLH and DEV for bringing builders together and creating opportunities to share projects, learn, and experiment! 🚀

If you're interested in knowledge handoffs, evidence-backed AI, or open-weight models, I'd love for you to check out RELAY and share your thoughts.

🔗 GitHub: https://github.com/Tanishaaaaaaa/relay-ai-handoff

🎥 Demo: https://youtu.be/5DhEMq-5jpc