:quality(80))
At The AI Conference Hack Day 2026, 23 teams built on BAND. In the projects that stood out, an agent or a human in the room checked the work before anything shipped.
On September 29, BAND was a core sponsor of The AI Conference Hack Day 2026, run by HackerSquad. 213 people checked in, and by the deadline teams had submitted 91 projects. 23 of them were built on BAND.
:quality(80))
The stage screen at Hack Day 2026.
What teams built on BAND
Most BAND projects shared one shape: a team of agents working in a BAND room, handing off work with @mentions, with either a critic agent or a human approval step before anything shipped. In the strongest ones, the check wasn't a box at the end of a pipeline. A critic agent read the work and sent it back with specific fixes. A few examples:
LaunchReel: five agents turn a website URL or app repo into a YouTube Short and a one-minute promo. One agent critiques every rendered frame. (GitHub)
Safe Scribe: a clinical handoff scribe where the agents argue before they commit, and no identifiable patient data leaves the room. (project, GitHub)
Thermal Crusoe: a technician uploads thermal evidence and a voice note. After a human approves the findings, BAND coordinates the repair across parts, scheduling and follow-up. (project, GitHub)
SYNTH: text one line after meeting someone, and agents in a BAND room research the person and draft the follow-up. (project, GitHub)
Dreamnetopi: a BAND room splits a shop owner's request across agents. Nothing goes live until the owner approves, and the room log is the receipt. (GitHub)
BAND projects did well in the wider judging too. Thermal Crusoe took Overall First Place and Safe Scribe took Overall Second Place. SYNTH won the Vultr prize, and CertAIn won Best Technical Implementation from Neo4j. The full list is on the event page.
The BAND prize: BAYBAY Crew
We picked BAYBAY Crew, built by Hao Dong Yuan, as the winner of the BAND prize. It also won Best Use of Neo4j.
The problem it targets is simple: travel chatbots invent events and get dates wrong. BAYBAY Crew plans a real day out in San Francisco, and every fact in the plan comes from a Neo4j graph built from the catalog of BAYLINK, a bilingual Bay Area events and guides site. The graph has 1,594 nodes and 3,053 relationships, including 267 dated events, 957 places and 161 transit stations.
Four agents work in one BAND room:
BAYBAY (host) takes the request from the web UI or an @mention, and delivers the final plan.
Scout turns the request into filters and runs Cypher queries for events, free-admission offers, nearby places and transit.
Planner builds a day of three to five stops using only Scout's candidates, with transit hints from shortest paths over the station graph.
Checker re-reads every stop from the graph (date, free or paid, minimum age, offer validity), and a second model judges the fit. If something is wrong, it sends the plan back to Planner with fixes, for up to three rounds.
Scout, Planner and Checker each run a different open model on Crusoe: DeepSeek-V4-Flash, DeepSeek-V4-Pro and Gemma 4 31B.
:quality(80))
The BAYBAY Crew app: the verified plan, the BAND room feed and Scout's candidates in the graph. Screenshot from the team's repo.
What stood out to us
The crew runs inside BAND. Every handoff is a BAND message with an @mention, and tool calls, results and agent thoughts are posted as room events. The agents only act on messages BAND delivers, so deleting the room stops the crew.
The critic can veto. In the team's demo run (about 30 seconds), the Checker rejects round one and @mentions the Planner with specific fixes. The Planner revises, and every stop in the final plan checks out against the graph. The team set a ceiling of three rounds. Inside it, the Checker decides whether the plan goes back, based on what it finds.
:quality(80))
Inside the BAND room: the Checker rejects round one. Screenshot from the team's repo.
It shows its work. The UI shows every Cypher query, the plan's subgraph, and the provider, model and latency of every model call. It also runs in Chinese and English.
Give builders a shared room and one day, and many of them reach for the same thing: a reviewer in the room, agent or human, whose job is to say no.
Thanks
Thanks to HackerSquad and The AI Conference for running the day, to the other sponsors, and to every team that built something. Special thanks to the 23 teams that built on BAND.
:quality(80))
The BAND booth at Hack Day 2026.
Links
BAYBAY Crew: project page, GitHub repo, demo video
All projects and winners: The AI Conference Hack Day 2026 on HackerSquad
BAND: band.ai
:quality(80))
:quality(80)){kind=small}
:quality(80))
:quality(80))
:quality(80)){kind=small}