Handoff

Here is your complete Devpost submission writeup ready to copy and paste:
## Inspiration
Group projects in university are notoriously chaotic, but generative AI has ironically made team collaboration more fragmented. Everyone on the team has their own favorite model: one person uses Claude 3.5 Sonnet for clear academic writing, someone else uses GPT-4o for math and coding, and another prefers Gemini. 

Because everyone works in their own private AI chat tabs, context gets shattered. Teammates waste 15 minutes repeatedly copying and pasting the syllabus, rubric, and previous drafts into their own chats. The final deliverable ends up being a disjointed "Franken-report" where Section 1 and Section 2 contradict each other. We wanted to build something like "Git commit and push," but designed specifically for collaborative human-AI teamwork.

## What it does
**Handoff** is a shared AI workspace for group projects where every teammate can use their model of choice, anchored to a single continuous project memory.

- **Model Freedom:** Teammates join the workspace, enter their name, and choose whichever model they prefer (Claude, GPT, Gemini).
- **The "Catch-Me-Up" Briefing:** Before you type a single prompt, your AI generates an automated briefing showing who worked last, what model they used, what was decided, and what the next task is.
- **Continuous Memory Extraction:** As you chat, key decisions, formatting constraints, and task ownership are automatically extracted and updated on the live Project Memory panel.
- **One-Click Session Handoff:** When finishing work, the user clicks **"End Session & Hand Off"**. The AI synthesizes the session into a structured note (`Done`, `Decided`, `In Progress`, `Next Task`) and commits it to the project timeline.
- **Seamless Continuation:** The next teammate opens Handoff on their laptop with a completely different model, says *"continue where Ibrahim left off,"* and their AI writes the next section using all previous team decisions without any re-prompting.

## How we built it
- **Backboard.io:** Powers the entire brain and memory layer. Each project maps to a Backboard Assistant that manages persistent long-term memory, document context, and threads across sessions. Backboard allows us to switch models seamlessly per message while preserving the shared memory state.
- **Frontend:** Built with Next.js, React, TypeScript, and Tailwind CSS. We designed a focused three-column workspace: a multi-model collaborative chat stream, a live Project Memory sidebar, and the Handoff Timeline.
- **Backend & APIs:** Next.js API routes that interface with the Backboard SDK/API to parse chat interactions, trigger real-time memory extractions, and compile structured handoff notes.

## Challenges we ran into
- **Cross-Model Context Normalization:** Different LLMs have distinct prompt structures and styling habits. Ensuring that a session started in Claude could be picked up by GPT without voice drift or hallucinated rules required strict memory structuring via Backboard.
- **Extracting Signal from Noise:** Teammates chat informally with AI. Filtering out trivial back-and-forth conversation while reliably capturing concrete project milestones (*"use IEEE citations"*, *"Sara owns slides"*) required tight extraction prompts.
- **Session Continuity:** Making sure asynchronous handoffs feel completely smooth between multiple browser windows and active sessions without state conflicts.

## Accomplishments that we're proud of
- **The "Catch-Me-Up" Moment:** The first time an AI greeted a teammate with an accurate, bulleted breakdown of what their partner did 10 minutes prior on a completely different model felt like genuine magic.
- **Zero-Friction Multi-Model Switching:** Watching Claude outline a research paper, hitting handoff, and having GPT-4o immediately pick up writing the methodology with full awareness of Claude's tone and constraints.
- **A Simple, Intuitive Ritual:** Creating the "End Session & Hand Off" button—giving group projects a clear punctuation mark that eliminates confusion on who is doing what.

## What we learned
- **Memory is the missing layer in collaborative AI:** Decoupling persistent memory from the specific LLM model is essential for true team workflows. Backboard's architecture proved that the model can be ephemeral as long as the memory is persistent.
- **Human workflows need structure:** Just giving people a shared chatbox isn't enough; introducing an explicit handoff ritual (like a Git commit) transforms how effectively teams collaborate asynchronously.

## What's next for Handoff
- **Live Collaborative Document Editor:** Embedding a Google Docs/Notion-style editor where the AI applies inline diffs and highlights changes directly tied to handoff notes.
- **Discord & Slack Webhook Notifications:** Broadcasting handoff summaries to the team's group chat as soon as a member finishes a work session.
- **Contradiction Detection:** Alerting team members in real-time if one teammate instructs the AI with a constraint that clashes with an earlier teammate's decision.