Inspiration
I work on a global Azure operations team and handle around 7–8 P1 incidents every week. The hardest P1s are rarely the technical ones. They are the ones at 3 AM, when you are half asleep, searching for your laptop, and trying to answer three questions: what is broken, how bad is it, and what changed. Afterwards, someone still has to rebuild the whole timeline from memory to write the RCA. I wanted a copilot that works the way an on-call engineer actually wakes up: hands-free and by voice.
What it does
NightShift is a voice‑first on‑call copilot built for the Alexa+ track.
Ask about incidents: “What P1s are active right now?” returns a short, spoken summary of active incidents and their severity. Check service health: “How is the payments service doing?” returns the current health status and the most likely cause based on recent changes. Log as you go: while you work, you can say “log that I restarted the node pool.” The copilot saves a timestamped entry to the incident timeline. Draft the RCA: when the incident ends, it turns the logged timeline into a structured RCA draft (impact, timeline, root cause, actions), so nobody has to rebuild it from memory the next day.
How we built it
The core is a Python MCP (Model Context Protocol) server that exposes three tools to Alexa+: get_active_incidents, get_service_health and log_and_draft_rca. Alexa+ decides when to call each tool based on what the user says, so the conversation stays natural. For this hackathon the incident and service data is simulated, and the demo shows the full voice flow end to end. The code is on GitHub, with a README explaining how to run it.
Challenges we ran into
- Designing answers for the ear instead of the screen. A good spoken incident summary has to be short and ordered by what matters most.
- Keeping the copilot honest. It reports only what is in the data and says clearly when it does not know the cause.
- Building and testing a working demo in a very short time.
Accomplishments that we're proud of
- Turned a real on-call pain point from my daily work into a voice-first copilot that covers the full incident loop: check status, log actions, draft the RCA.
- Got a working end-to-end voice demo with three MCP tools running together.
- Wrote spoken answers that are short, ordered by severity, and honest about what is unknown.
What we learned
Voice is a surprisingly good interface for operations work, because the moments when you need help are exactly the moments when your hands and eyes are busy. I also learned that much of an Alexa+ experience depends on writing clear tool descriptions, so the assistant knows when to call each one.
What's next for NightShift: Alexa+ On-Call Copilot
- Connect to real sources such as Azure Service Health and ServiceNow instead of simulated data.
- Add voice escalation, for example "page the database on-call".
- Add a confirmation step before any action that changes a system.
- Support multiple languages so teams across regions can use it.
Log in or sign up for Devpost to join the conversation.