Four AI agents coordinate in a Band room to plan a real day out in San Francisco from a Neo4j knowledge graph of BAYLINK's verified local catalog, reasoning with open models on Crusoe. A Checker agent re-verifies every stop against the graph (and can veto the plan) before a human sees it. WHY: travel chatbots invent events and get dates wrong. The crew splits the job so every fact comes from the graph and every plan is checked. AGENTS (every handoff is a Band message with @mentions; tool calls, results and thoughts are Band room events; agents act only on messages Band delivers, so deleting the room stops the crew): - BAYBAY (host): takes the request (web UI or @mention in the Band room), delivers the verified plan. - Scout: request -> filters (Crusoe: DeepSeek-V4-Flash) -> Cypher on Neo4j Aura: events on that date, free-admission offers, places nearby, transit. - Planner: a 3-5 stop day ONLY from Scout's candidates (Crusoe: DeepSeek-V4-Pro); transit hints = shortestPath over the graph's station network. - Checker: re-reads every stop from Neo4j (date, free/paid, minimum age, offer validity) + a different model (Crusoe: Gemma 4 31B) judges the fit; wrong -> @Planner with fixes (up to 3 rounds); right -> @BAYBAY. Dependent handoffs + a critic that can veto. NEO4J: 1,594 nodes / 3,053 relationships from BAYLINK (baylink.us): 267 dated Bay Area events, 18 OSM-checked venues, 957 places, 72 neighbourhoods, 161 stations on 7 lines, verified free-admission offers, new openings. The UI shows every Cypher query, the plan's subgraph and a dashboard of three graph questions. CRUSOE: every reasoning step runs on Crusoe Managed Inference, one open model per agent; the UI shows provider, model and latency on every call. DEMO: a real run (~30 s): Checker vetoes round 1 in the Band room, Planner fixes it, all stops verified, with links to real BAYLINK pages. Bilingual (Chinese / English). Demo video: demo/baybay-crew-demo.mp4 in the repo.
Built at The AI Conference Hack Day 2026
BAYBAY Crew: real San Francisco days, planned by a crew of agents
Four AI agents coordinate in a Band room to plan a real day out in San Francisco from a Neo4j knowledge graph of BAYLINK's verified local catalog, reasoning with open models on Crusoe. A Checker agent re-verifies every stop against the graph (and can veto the plan) before a human sees it. WHY: travel chatbots invent events and get dates wrong. The crew splits the job so every fact comes from the graph and every plan is checked. AGENTS (every handoff is a Band message with @mentions; tool calls, results and thoughts are Band room events; agents act only on messages Band delivers, so deleting the room stops the crew): - BAYBAY (host): takes the request (web UI or @mention in the Band room), delivers the verified plan. - Scout: request -> filters (Crusoe: DeepSeek-V4-Flash) -> Cypher on Neo4j Aura: events on that date, free-admission offers, places nearby, transit. - Planner: a 3-5 stop day ONLY from Scout's candidates (Crusoe: DeepSeek-V4-Pro); transit hints = shortestPath over the graph's station network. - Checker: re-reads every stop from Neo4j (date, free/paid, minimum age, offer validity) + a different model (Crusoe: Gemma 4 31B) judges the fit; wrong -> @Planner with fixes (up to 3 rounds); right -> @BAYBAY. Dependent handoffs + a critic that can veto. NEO4J: 1,594 nodes / 3,053 relationships from BAYLINK (baylink.us): 267 dated Bay Area events, 18 OSM-checked venues, 957 places, 72 neighbourhoods, 161 stations on 7 lines, verified free-admission offers, new openings. The UI shows every Cypher query, the plan's subgraph and a dashboard of three graph questions. CRUSOE: every reasoning step runs on Crusoe Managed Inference, one open model per agent; the UI shows provider, model and latency on every call. DEMO: a real run (~30 s): Checker vetoes round 1 in the Band room, Planner fixes it, all stops verified, with links to real BAYLINK pages. Bilingual (Chinese / English). Demo video: demo/baybay-crew-demo.mp4 in the repo.
Keep exploring what builders shipped.
HackerSquad project
RunIt
RunIt is the AI chief of staff I run my business on. It reads my texts, email, Instagram DMs, calendar and CRM, and drafts every reply in my own voice, with the context already in it. Plaud is what makes that context complete: most of what matters happens in person or on a call, and it never lands in a text thread or an inbox. Plaud captures it, and RunIt turns it into context. What the demo shows, live on my phone with my real data: 1. Connectors: Plaud, Messages, Gmail, Google Calendar, Instagram, Notion and phone calls, all feeding one context layer. 2. Pre-drafted replies: every text, email and DM that comes in already has a reply drafted from everything RunIt knows about that person, including what we said in a Plaud-recorded conversation. One tap sends it from my own iMessage or Gmail. Each draft shows its sources. 3. Action bundles: when a reply depends on something (check the calendar, confirm a detail), RunIt stages those actions first and waits for my approval. I can send with notes or redraft with notes. 4. Auto: RunIt can carry a conversation for me toward a goal I set (who, what, how long), and only pulls me in when it's stuck. 5. "What did I learn today?": RunIt answers from today's Plaud recordings of the conference talks, with the key takeaways, and offers to send them to someone. Built today at Hack Day: a Plaud Embedded SDK iPhone app (Capacitor) that binds a Plaud NotePin S or Note Pro, syncs recordings, gets a speaker-labeled transcript from the Plaud Transcription API, and hands it to RunIt. RunIt never shows a transcript. It shows what you owe people (drafted in your voice), what they owe you (tracked, with the day it will check in), and what's worth remembering about them (saved only when you say so, brought back at the right moment). Plus a Plaud connector in Settings and "recorded today" context in the chat. Nothing acts on its own. Every send waits for one tap, passes an allowlist and a rate cap, and leaves a receipt. Plaud hears it. RunIt makes sure it gets done.
HackerSquad project
Voice Agent Metaharness: one ecosystem for all of your agents
One ecosystem for all of your agents: one Claude across every surface (phone voice, web, Claude Code, an agent team, a WebXR world) sharing one picture of your day. Plaud: a clip-on recorder captures what happens around you. contextlog: a shared-context MCP server hands that day to whichever surface you return to (while_away), lets every agent share what's live in its context (context_ping), and catches ideas you say out loud (idea_log). Band.ai: an org of 9 agents (PM, Architect, Frontend, Backend, QA, DevOps, UXR, Research) led by Claude Code; each logged idea becomes a GitHub issue and project card, posted to the Band room where the PM agent triages it. Similarweb: UXR and market research (41 API calls, 16 domains): ambient capture is growing fast (Plaud +155% YoY, Granola +257%) while recorders commoditize, and Plaud's audience already overlaps with Claude's, so the value is the shared context above the device. Also registered in DuploCloud as a REST provider. DuploCloud DevKit: the control plane, with Neo4j (mcp-neo4j-cypher), contextlog, Band and Similarweb registered as providers, scopes and MCP servers. Dreamspace: the spatial surface; its guide Lumen speaks with spatial audio and voice-codes the world on request, and Claude can join hands-free over MCP. Vultr: hosts the shared MCP server over HTTPS. and github.com/rachael/contextlog Repos: github.com/rachael/dreamspace and github.com/rachael/duplocloud-setup
HackerSquad project
TheraApp
Structured CBT for mild to moderate anxiety