Inspiration
What it does
How we built it
Challenges we ran into
Accomplishments that we're proud of
What we learned
What's next for Opening Night
Inspiration
Anyone who has worked a live show knows the feeling. It is 3pm, doors are at 7, and two things break at once. The plan lives in one person's head, the money lives in someone else's inbox, and the crew finds out by rumor. Small venues and touring acts run on spreadsheets and group chats, and one bad afternoon can cost them the whole night.
I wanted to build the tool I would want in that moment. Something that sees the problem, works out what can still be saved, pays the right people, refunds the right fans, and does not say "go" until everyone who matters has confirmed.
What it does
Opening Night is an operating system for the day of a show. The demo follows a concert at Alder Hall on a very bad afternoon.
- A delayed flight and a dead lighting console hit the schedule. The whole night reflows on a live Gantt, crew scheduler and calendar, and you can see exactly what slipped and why.
- AI proposes recovery plans. A scheduling engine checks each one against real constraints like crew rest, venue curfew and load in times, and throws out the ones that do not hold up.
- A replacement console is sourced from three suppliers. PayPal authorizes $780, captures the $480 that was actually used, and releases the $300 hold.
- 80 ticket holders are charged through saved PayPal wallets for $3,370. The 38 VIP guests whose meet and greet was cut get $15 back each, straight to the original payment.
- All 11 crew members get an updated run sheet on their phone and have to acknowledge the changes. In the demo you scan a QR code and become the eleventh crew member yourself.
- Eight readiness gates turn green one at a time. Only then can the producer press GO.
- After the show, crew are paid with PayPal Payouts and suppliers are settled with PayPal invoices.
How we built it
- Next.js on Vercel with Postgres for everything that has to survive a refresh, a retry or a crash.
- Bryntum Gantt, Scheduler and Calendar for the planning surfaces, with the Bryntum engine running on the server so every AI suggestion is checked by the same engine the user sees.
- PayPal Orders (authorize, capture, void), Vault for saved wallets, Refunds, Payouts, Invoicing and Webhooks, called through the PayPal server SDK generated by APIMatic.
- The PayPal Agent Toolkit gives the wrap up assistant read access to real order and invoice records.
- OpenAI models read source documents, propose recovery strategies, and power a voice assistant inside the schedule.
- Every payment has a durable operation identity, so a lost response or a double click can never charge someone twice.
How AI uses PayPal
AI never moves money on its own. It proposes a plan and the exact payment instruction that goes with it. A person approves it. The system then executes it against PayPal and reads the result back from PayPal before showing it as done. The assistant can look up orders and invoices through the Agent Toolkit to answer questions about where the money went.
Tools used and how
- PayPal: Orders, Vault, Refunds, Payouts, Invoicing, Webhooks, Agent Toolkit (sandbox)
- APIMatic: generated PayPal SDK and API context used to build and verify the payment layer
- Bryntum: Gantt, Scheduler, Calendar and the server side scheduling engine
- OpenAI: document reading, recovery planning, voice assistant
- Vercel and Postgres: hosting, cron jobs and durable storage
Challenges we ran into
The hard part was not calling PayPal. It was being honest about state. A saved wallet is not reserved money. An authorization is not a capture. A webhook can arrive twice or not at all. We rebuilt the payment layer around one rule: nothing is shown as paid until PayPal says so.
The second challenge was the crew phone step. A show cannot start because a dashboard says so. It starts when people confirm. We made that a real gate and built a QR flow so a judge can feel it.
Accomplishments that we're proud of
- A full night, from disaster to wrap, runs in under a minute in the guided tour.
- A real sandbox run of the whole money story: 80 captures for $3,370, 38 refunds for $570, and the $780 / $480 / $300 rental lifecycle.
- AI plans are validated by a real scheduling engine instead of being trusted.
What we learned
Payments are a state machine, not a button. And the best use of AI in operations is to propose and explain, then let a strict engine and a human decide.
What's next
Multi city tours, real supplier onboarding, and learning from past shows so the next recovery plan is better than the last one.
Notes for judges
The guided tour replays recorded PayPal outcomes so it is instant and free to run, and it says so on screen. The same code paths run against the live PayPal sandbox, and the README lists what was verified there.
