Developer discovery engine · v0.1

Find developers who build like you

GitHub Social Agent discovers developers with similar technologies, projects, and interests — helping you build a meaningful technical network.

View on GitHub How it works
0–100 similarity score shadow-mode ML local-first SQLite self-growing graph
The problem

GitHub has millions of developers.
Finding the right ones is hard.

Popular developers are not always relevant developers. The most valuable connections are people who share your technologies, engineering interests, and project mindset.

Millions of developers

Every profile, repository, and follower edge is a candidate — and almost all of it is noise for you.

Hard to find relevant people

Searching by keywords or popularity rarely surfaces engineers who actually work on the same problems as you.

Manual discovery does not scale

Reading hundreds of profiles by hand takes hours — and the results are never comparable.

The solution

Your personal GitHub discovery agent

The agent continuously analyses the GitHub ecosystem around your profile and ranks developers by technical similarity.

Discover

Traverses follower networks and discovers new developers.

follower graph

Analyse

Collects repositories, languages, topics, and activity data.

repo intelligence

Score

Ranks developers using your own GitHub profile as a baseline.

0–100 ranking

Connect

Can automatically follow relevant developers within configured limits.

automated networking
How it works

From GitHub data to developer connections

01

GitHub API

Collector and discovery workers read profile, repository, and follower data within rate limits.

api.github.com
02

Repository Collection

Languages, topics, and activity recency are stored for every developer.

languages · topics
03

Developer Profile

Every discovered user becomes a node in a local similarity graph.

SQLite · local-first
04

Similarity Scoring

Profiles are compared against your baseline using histogram intersection and Jaccard similarity.

score 0–100
05

Ranked Developers

Top matches are surfaced — and can be followed within configured daily limits.

DAILY_FOLLOW_LIMIT
Interactive demo

See how developer similarity works

Your profile is compared with a developer candidate. Adjust the candidate's stack and watch the similarity score react in real time.

Your profile

owner baseline · auto-synced
Java45%
Python25%
JavaScript15%
Other15%
backend spring docker kubernetes

Developer candidate

adjust languages and topics
Java50%
Python20%
Go10%
Other 20%
Topics
0/ 100
Similarity score
higher = closer to your stack
Language similarityhistogram overlap of shared stacks0 / 30
Topic similarityshared topics · related topics half-credit0 / 20
Repository activitysimulated recency signal0 / 15
Profile qualitybio · repo count · followers0 / 35
What drives the score

This is a demo of the scoring principle — not a real analysis of your GitHub profile. The live system uses histogram intersection (languages), Jaccard similarity (topics) and repo recency (activity).

Scoring

Recommendations based on engineering signals

The score uses measurable technical signals instead of popularity alone.

0/ 100
Developer Score
Languages 30 Topics 20 Activity 15 Quality 35
Profile qualitybio · repo count · followers35 pts
Languageshistogram intersection with your stack30 pts
TopicsJaccard similarity of topic sets20 pts
Activitydays since last repository update15 pts

The system compares programming languages, repository topics, and activity patterns between developers — not follower counts or popularity.

Machine learning

Experimental machine learning layer

The project includes an ML model that learns from historical follow decisions and evaluates prediction quality on real GitHub data.

The model currently runs in shadow mode — its predictions are logged and evaluated, but never used to make decisions.

SHADOW MODE · predictions logged, not applied

Production decisions are still controlled by the deterministic scoring system.

1

Historical Data

Past follow decisions and profile signals become training examples.

2

ML Training

A PyTorch model learns which signals predict a good connection.

3

Prediction Evaluation

Precision and recall are measured against real GitHub outcomes.

4

Quality Analysis

Metrics decide whether the model is ever promoted out of shadow mode.

Engineering

Built for long-running operation

A background service designed to run for weeks without supervision.

Self-growing discovery

Continuously discovers new developers through follower networks.

DISCOVERY_RATE_LIMIT_PER_HOUR

Rate-limit aware

Designed around GitHub API constraints with retries and cooldowns.

401/403 retry · 2h cooldown

Local-first

Stores collected data in your own SQLite database.

data/github_social.db

Docker ready

Runs as a standalone background service.

Dockerfile · compose

Dashboard

Provides a web interface for exploring collected data.

FastAPI · React

Graceful operation

Supports safe shutdown and persistent progress.

SIGINT · SIGTERM
Architecture

How the system is wired together

Workers collect data, the scoring engine ranks it, and the ML model watches from shadow mode.

GitHub API REST · rate-limited Discovery Engine traverses follower graph Repository Collector repos · languages · topics Graph Discovery follower network growth Scoring Engine 0–100 similarity ranking ML Shadow Model predictions logged · not applied Dashboard explore collected data

data flow: GitHub API → Discovery Engine → Scoring Engine → ML Shadow Model → Dashboard

Build your developer network automatically

Discover engineers who share your technical interests without manually searching through thousands of profiles.

Open GitHub Repository
$ python main.py --silent