Takoa

• 

  This is Takoa

• 

  Benchmarks to ensure is running

• 

  Talk to Takoa

• 

  Only visible to devs + judges

Inspiration

Friend-finder apps promised real connection, but the loneliness epidemic kept growing. Swipes, ghosting, and endless choice make friendships feel disposable. We wanted something different: a friendship platform that learns your context through conversation, then forms a small, cohesive group you can actually meet. Takoa is our push against one-size-fits-all social apps. It is built to get you real connections, not keep you stuck in a loop.

What it does

Takoa helps you find your people without swiping or quizzes. You chat naturally with a guided onboarding assistant. The system extracts conversational signals into a structured profile (traits, interests, confidence) - no AI here, semantic logic + OAI embeddings' are used. A live 3D social space shows how people sit in the similarity graph. You are placed into a circle of five with shared context, not ranked by AI. Once confidence is high enough, a simple group plan is revealed to kickstart a real meetup (currently limited to university grounds for safety)

How we built it

Next.js 14 (App Router) with React + TypeScript.. Conversational signal extraction from chat into a structured profile. Vector representation with L2 normalization and cosine similarity. kNN retrieval for candidate neighbors. 3D visualization using react-force-graph-3d (force and embedding views). DBSCAN clustering and UMAP-style embedding layout ( the graph in the picture).

Challenges we ran into

The prompts - friends and AI aren't a fan favourite idea, rightfully so. We really had to make sure our application didn't all of a sudden turn into a AI companion, therapist or anything else other than help us parse information from your context. Architecturing this, the backend is really heavy but at hackathons we are showing off the exciting part! so we needed to really make sure we were creating a technical enough project worth our caliber and time while also remaining a "commerically" fun project. Balancing explainability with polish in the 3D space, inspired by a quantum workshop I went to, except visualising it in your head is easier than visualising it via code. This ML stuff, we were reading the research papers and no amounts of AI was simplifying some of these concepts lol, we got a pretty good sense of the architecture by the end of the project and my notebook is FULL of notes. Designing a plan trigger that is deterministic and demo-safe. Keeping the system human-centered while still using strong technical signals, we genuinely do think of this as a project that has not yet been implemented at a large scale yet so we were really building this one without a real reference.

Accomplishments that we're proud of

A visual, interactive social graph that updates as you chat. A clean separation between AI signal extraction and deterministic matching logic. A product story that prioritizes real connection over engagement metrics. This entire thing! Scoured the internet for which technolgy is better, which algorithm is best for the amount of data we have, clustering algo, density based or centroid, visualising embeddings and whatnot. We spent 5 hours of the start of this hack staring at research papers/websites/perplexity trying to explain the concepts.

What we learned

People trust systems more when the logic is visible and consistent. Architecting is really overwhelming. Design engineering this is hard, a lot of the data we attained from our backend wasn't translated as well as it could have been, i'd blame time but it also is just tricky. Great matchmaking starts with great questions, not better swipes - tbf we knew this but we got surer of it as we built up this project. People really really really reallly don't like it when AI tries to become a bestfriend/companion (lots of news articles about this, rightfully).

What's next for Takoa

We managed to implement every single sub idea for this project actaully. Embeddings, semantic logic, deterministic ML/math, have the confidence thresholds we wanted, UI logic, graph visualisation and so much more.

If we do like our idea a lot, our next goal would probably be changing up the architecture to fit more than 100 people. Our parameter tuning resulted in their best results currently, with added data and actual real data, chances are the value of our parameters, vectors, etc, also changed drastically.

Built With

• lots
• more
• next.js
• openai
• typescript