AI that helps people meet, not decide for them.
Team Spark takes two people who are already interested in each other and helps them have a better first conversation and a shared plan. Built for HackTX 2026, Meta's “Bringing people closer together with AI” track.
The problem: the first hello
Most hackathon teams form in a hurry on a chat server, between people who already know each other. If you are new, shy, from a smaller school, or just don't have a network, the cold start is the hardest part. You scroll profiles, never send the message, or send “hey” and hear nothing. Great teammates never meet, and some people give up before the event starts.
Team Spark is built for them. It lowers the cost of saying hello, and it keeps the decision to team up entirely with the people involved.
Three moments where Muse Spark helps
1. Explore this match
You choose to explore someone. Muse Spark reads only what you both already share and writes why you might complement each other, one real thing you have in common, a friendly first message in your voice, and a question that shows whether you want to build the same kind of thing.
2. A short shared session
After you connect, you each answer three prompts: the problem you care about, what you want to contribute, and what makes a good teammate. Neither of you sees the other's answers until you have both answered, so nobody copies anybody.
3. A Build Brief you can edit
Muse Spark turns both answers into a project direction you both support, who it helps, roles taken from what each of you said, agreements about talking and being heard, and a 30-minute first-step plan. You edit every word before sharing it. Then one click makes the team and invites the other person, who still has to accept.
What the AI sees, and what it never sees
Before anything is sent, the page shows you the exact list below, filled in with your real fields. You press the button to say yes. Nothing from Explore is stored.
Sees
- Headline, bio, skills, interests and the roles you want
- Experience level and what you are looking for
- How you like to work and when you are free
- School and major, only if you show them to the other person
- The three answers you write in a Team Spark session
Never sees
- Your private messages
- Email, phone or any contact detail
- Links, resume or photo
- Anything you hid from the other person
- Anyone else's profile
Responsible by design
Opt in, every time
Nothing runs until a person presses the button after seeing what will be sent.
No scores, no rankings
Muse Spark is told never to rank people or call anyone the most compatible. It says so plainly when two people have little in common.
A human edits first
The brief is a draft only its author sees. They change anything, then share it.
People decide
AI suggests; people connect, accept invitations and join teams. Nobody is added to a team without saying yes.
Organizers see counts only
Totals such as briefs shared and teams created. Never answers, briefs or names.
Private messages stay private
They are never sent to the model, and organizers can't read them.
Social good: a warmer way into a team
The people who benefit most are the ones the usual process leaves out: first-timers, quieter people, students from smaller schools, anyone outside the circles that already form teams together. The facilitator is instructed to be welcoming: plain language, no gatekeeping, no assumptions about experience or background, and to suggest ways to learn together instead of talking down.
Every Build Brief includes who the project could help and an agreement about making sure both voices are heard, so teams start with kindness and a purpose beyond the prize. Organizers can see in aggregate whether connections are turning into teams, so they can reach out to people who haven't found one.
Under the hood
- Model: Muse Spark (
muse-spark-1.1) through the Meta Model API, called from the server only. The key never reaches the browser. - Structured output: every reply must be a JSON object that passes validation, or the person sees a friendly error instead of raw AI text.
- Limits: rate-limited per person, with logging of model and token counts only, never prompts or answers.
- Platform: HackMatch (Next.js, Postgres), the teammate-finding app that existed before the event. Team Spark was built during the submission window; see the repository for exactly what.