Constellation

Constellation is an AI-powered event networking platform that helps you find the right person in the room, understand why you should talk to them, and turn that conversation into a real connection.

Inspiration

Hackathons, conferences, career fairs, and networking events put hundreds or even thousands of interesting people in the same place. The problem is that being surrounded by the right people does not mean you will actually meet them.

You might walk past someone building the exact technology you want to learn about, a recruiter looking for your skill set, or another student with the missing piece for your next project and never know it.

Traditional networking platforms help you maintain connections after you meet someone. We wanted to solve the harder problem:

Who should I talk to right now, and why?

That idea became Constellation.

Instead of making users endlessly search profiles, Constellation combines their interests, skills, goals, and real-world proximity to surface the people they are most likely to have a meaningful conversation with.

What it does

Users build a profile using information such as their resume, GitHub activity, interests, and what they are currently looking to learn, build, or offer.

Our AI converts that information into structured interests across areas such as technical interests, career goals, academics, and personal interests.

When a user turns on Open to Meet, Constellation:

Finds compatible people nearby without publicly exposing exact locations.
Ranks potential connections using shared interests, complementary goals, and more specific signals such as niche interests.
Shows a match score and explains why the two people should talk.
Generates personalized conversation starters based on actual common interests.
Uses Bluetooth proximity to estimate whether a match is nearby.
Lets users privately indicate that they want to meet. A rejection is never shown to the other person.
Unlocks messaging and temporary meeting tools once interest is mutual.
Helps both people find each other using rough proximity instead of permanently sharing precise location.
Detects that the two people were physically near each other and lets them confirm what they discussed.
Creates a connection only after both people choose to connect.

We also built an AI assistant that can answer questions such as:

"Who here should I talk to about RAG?"

or

"What should I ask this person?"

Instead of treating networking as collecting as many contacts as possible, Constellation focuses on helping people create fewer, more meaningful connections.

How we built it

Our mobile application is built with React Native, Expo, and TypeScript. We use Supabase for authentication, database infrastructure, realtime data, and account management.

The AI and machine learning system is served through a Python FastAPI backend.

Our matching pipeline starts by converting messy profile information into structured interests. We use Claude for structured interest extraction and grounded explanations, then use sentence embeddings to represent those interests mathematically.

Rather than treating matching as a simple classification problem, we designed it as a ranking problem. For every user, the system tries to rank everyone else by how valuable a conversation between them could be.

Some of the signals include:

Similarity across technical, career, personal, and academic interests
Shared niche interests
What one person is seeking compared with what another person can offer
Cross-disciplinary "bridge" connections
Previous interaction outcomes

We built both an explainable weighted ranking system and a learned ranking pipeline using models including logistic regression and LightGBM. In our synthetic evaluation pipeline, the learned model improved over our original hand-tuned matcher, although these results are based on simulated outcomes rather than real production users.

For physical discovery, we integrated Bluetooth Low Energy. BLE allows Constellation to reason about rough proximity without turning the app into a live location tracker.

We also built an encounter model that examines Bluetooth signal patterns over time. Because Bluetooth alone cannot reliably prove that two people actually spoke, we intentionally do not allow it to create a connection automatically. Both users still have to confirm the interaction.

That combination of AI, ML, mobile development, realtime infrastructure, and physical proximity is what makes Constellation more than a recommendation feed.

Challenges we faced

Making AI matching actually meaningful

It is easy to say two people both know Python. That does not necessarily mean they should meet.

We had to think much more carefully about what predicts a useful conversation. A rare shared interest can be far more important than a common skill, and sometimes the strongest connection is complementary rather than similar.

Someone looking to learn about computer vision may be a better match with someone experienced in computer vision than with someone who simply has the same background.

Connecting the digital and physical worlds

Finding someone in a database is easy. Determining whether that person is actually nearby is much harder.

Bluetooth signals are noisy. Walls, phone orientation, movement, and crowded rooms can all affect RSSI measurements. Someone sitting at the same table can look similar to someone actively having a conversation.

This forced us to design Bluetooth as supporting evidence rather than absolute truth and build human confirmation directly into the product.

Privacy

A networking app involving proximity can become invasive very quickly.

We designed Constellation so users explicitly choose when they are open to meeting people. Exact location is not exposed during general discovery, mutual interest is required before additional meeting tools unlock, and neither person sees when the other rejects a match.

We wanted the system to help create serendipity without making users feel tracked.

Integrating everything

Constellation touches authentication, mobile interfaces, realtime databases, APIs, LLMs, embeddings, ranking models, Bluetooth, location permissions, messaging, and connection graphs.

Getting each component working independently was only half the challenge. The harder part was making the complete experience feel like one continuous flow:

discover → understand → mutually match → meet → talk → verify → connect

What we learned

The biggest thing we learned is that recommendation systems are not just about maximizing similarity.

Good human connections often come from a combination of similarity, complementarity, timing, proximity, and intent.

We also learned how different building an AI feature is from simply calling an AI API. The model output has to become structured data, that data has to interact with deterministic ranking logic, and the final recommendation has to be explainable enough that a real person trusts it.

On the engineering side, we gained experience building across a large stack: React Native, Expo, Supabase, FastAPI, machine learning models, LLM integrations, embeddings, Bluetooth, realtime systems, and mobile permissions.

Most importantly, we learned that the best technology in this project is technology that eventually gets out of the way. The goal is not for someone to spend more time looking at Constellation.

The goal is for them to put their phone down and talk to someone they otherwise never would have met.

What's next

We would love to deploy Constellation at larger hackathons, conferences, universities, and career fairs and train the ranking system on real interaction outcomes.

With enough real-world data, the system could learn which kinds of introductions actually lead to meaningful conversations, identify communities that rarely interact, and surface valuable connections across those boundaries.

Our long-term vision is simple:

In a room full of strangers, Constellation should help you find the few people who could change what you build, learn, or do next.