Overview

Devfolio Analyzer is a Next.js application that combines GitHub's REST API with an LLM-based scoring pipeline to produce structured, actionable evaluations of developer profiles.

The core insight is simple: GitHub profiles contain a lot of signal, but that signal is buried in raw data. Recruiters don't have time to manually audit 40 repos, and developers rarely know how their profile reads to someone who isn't them. Devfolio Analyzer bridges that gap.

How it works (high level)

GitHub Username
     │
     ▼
GitHub REST API (user, repos, events, contributions)
     │
     ▼
Data Normalization Layer (scoring-ready structured data)
     │
     ▼
LLM Analysis (prompt-engineered, structured output)
     │
     ▼
Score + Feedback JSON
     │
     ▼
Frontend UI (Next.js — score cards, feedback list, charts)

The process from username submission to result delivery takes approximately 4–8 seconds depending on repo count and LLM response time.

Scoring model

The total score is a weighted composite across five dimensions:

Dimension	Weight	What it measures
Project Quality	30%	README, structure, tests, CI/CD, meaningful commit history
Tech Stack	20%	Language diversity, recency, framework maturity
Contribution Consistency	25%	Activity frequency, gaps, streak length
Portfolio Presentation	15%	Profile README, pinned repos, bio quality, links
Community Engagement	10%	External PRs, issue participation, forks to meaningful repos

Each dimension produces a sub-score from 0–100. The composite score is the weighted sum.

Score ranges

Score	Label	Interpretation
85–100	Excellent	Strong signal. Profile is polished and active.
70–84	Good	Solid but has specific gaps worth addressing.
50–69	Fair	Active developer but presentation or consistency needs work.
30–49	Weak	Sparse activity or underdeveloped projects.
0–29	Poor	Minimal public signal. Hard to evaluate technically.

What the LLM actually does

The raw GitHub data is too noisy to feed directly into an LLM context window. Before the LLM sees anything, the data normalization layer reduces it to a structured summary:

Top 6 repos by stars + recency
Aggregate language stats (% breakdown)
Contribution streak metadata
Presence flags (has README, has tests, has CI config, has pinned repos)

This normalized object is injected into a prompt template that instructs the model to:

Evaluate each scoring dimension independently
Return a JSON-formatted response (no markdown, no prose outside the schema)
Produce specific, non-generic feedback per dimension

The model is explicitly instructed not to award points for quantity. Ten low-effort repos score worse than three well-maintained ones.

Design decisions

Why not use GraphQL? GitHub's REST API v3 is more stable across rate limit scenarios and easier to cache at the edge. The GraphQL API requires more careful query scoping and has stricter rate limits for unauthenticated requests. For the data points Devfolio Analyzer needs, REST is sufficient.

Why LLM for scoring instead of a deterministic algorithm? The qualitative dimensions — "is this README actually useful?" — don't reduce cleanly to boolean checks. An LLM can read a README and evaluate it the same way a human would. The quantitative dimensions (contribution frequency, language counts) are computed deterministically and passed as context.

Why Next.js API routes instead of a separate backend? Deployment simplicity. Everything runs on Vercel — no separate server to manage, no CORS issues in development, and the API routes are close to the frontend. If throughput demands grow, the API layer can be extracted to a standalone service without changing the frontend.