GitDeep

An AI-powered GitHub profile analyzer that delivers brutal, honest assessments of any developer's work — from code quality to career trajectory.

🚀 Try GitDeep Live →

GitDeep is a client-side web application that uses advanced AI models to evaluate GitHub profiles with unprecedented depth. Unlike traditional portfolio scanners, GitDeep performs behavioral analysis, buzzword detection, career slope tracking, and per-repository assessments to give you the full picture.

---

✨ Features

🎯 Dual Assessment Modes

• Employer Mode: Brutal hirability analysis with no sugar-coating. Evaluates weaknesses, red flags, buzzword density, and provides a 1-10 hirability score with role recommendations.
• Developer Mode: Same honest assessment plus a detailed mentorship plan with actionable improvement steps, project suggestions, and learning paths.

🔍 Deep Analysis Capabilities

• Career Slope Detection: Tracks trajectory from account creation to present — Rising Star, Steady Maintainer, Declining Activity, or Sporadic patterns. Includes burnout risk assessment.
• Buzzword vs Reality Check: Compares bio/README hype words against actual code. Detects when developers claim "AI/ML expert" but write basic CRUD apps.
• AI Usage Quality Assessment: Distinguishes lazy AI slop (generic gradients, broken links, emoji spam) from high-quality AI orchestration (production-grade architecture, intentional tech choices).
• Behavioral Analysis: Confidence vs Arrogance scoring. Detects toxic patterns in PRs and communication style.
• Per-Repository Scoring: Each repo gets an independent 1-10 score with verdict, key highlights, and red flags.
• SWOT Analysis: Strengths, Weaknesses, Opportunities (market positioning), and Threats.
• Merged PR Analysis: Evaluates contributions to external repositories, not just owned repos.

🔒 Privacy-First Architecture

• No Database: 100% session-based. Zero data storage on servers.
• Client-Side Processing: All AI calls happen from your browser. Your API keys never touch our servers.
• No Tracking: No analytics, no cookies, no telemetry.

🤖 AI Provider Flexibility

Choose from 12+ AI providers:

• Cloud: Gemini (free tier), OpenAI, Anthropic, Groq, DeepSeek, Mistral, OpenRouter, Grok, Qwen, Kimi, NVIDIA NIM
• Local: Ollama (run models like Llama 3.1, Qwen 2.5, Mistral locally)

📊 Rich Visualizations

• Language distribution pie charts
• Strength/weakness radar charts
• Career timeline with growth meter
• Comparative candidate analysis (Employer Mode)

🔧 Advanced Features

• GitHub PAT Support: Analyze private repos and increase API rate limits
• Custom Questions: Ask the AI specific questions about a candidate's suitability
• Candidate Comparison: Compare up to 5 developers side-by-side with AI-powered ranking
• Prompt Size Options: Full (1200 tokens) for cloud models, Small (400 tokens) for weak local models

---

🚀 Quick Start

Prerequisites

• Node.js 18+ and npm
• An AI provider API key (Gemini recommended — has free tier)

Installation

# Clone the repository
git clone https://github.com/Yuvraj-Sarathe/GitDeep.git
cd GitDeep

# Install dependencies
npm install

# Run development server
npm run dev

Open http://localhost:3000 in your browser.

Configuration

1. Get an API Key (Recommended: Gemini)

   • Visit Google AI Studio
   • Click "Get API Key" → Create API key (free tier available)
   • Copy the key

2. Configure in App

   • Click the Settings gear icon (bottom-right or top navigation)
   • Select AI Provider: Gemini API
   • Paste your API key
   • Model: gemini-2.5-flash (default)
   • Prompt Size: Full (recommended)

3. Optional: GitHub PAT Token

   • Go to GitHub → Settings → Developer settings → Personal access tokens → Fine-grained tokens
   • Generate token with repo and user scopes
   • Add in Settings → GitHub PAT Token field
   • Benefits: Higher API rate limits, private repo analysis, more PR data

Usage

1. Enter a GitHub username (e.g., torvalds)
2. Choose mode: Employer (hiring) or Developer (mentorship)
3. Click Generate Assessment
4. Wait 10-30 seconds for analysis
5. Review the detailed report, metrics, and recommendations

---

🐳 Docker Deployment

# Build the image
docker build -t gitdeep .

# Run the container
docker run -p 3000:3000 gitdeep

Access at http://localhost:3000

---

🧠 Using Local AI Models (Ollama)

For complete privacy and no API costs, run AI models locally:

Setup Ollama

1. Install Ollama

   • Download from ollama.com
   • Install for your OS (Windows/Mac/Linux)

2. Pull a Model

   # Recommended for 8GB RAM (best balance)
   ollama pull qwen2.5:7b
   
   # Alternative: Mistral 7B
   ollama pull mistral
   
   # For 16GB+ RAM (better quality)
   ollama pull llama3.1:8b
   
   # ⚠️ NOT recommended (too small for this task)
   # ollama pull phi
   # ollama pull llama3.2:1b

3. Start Ollama Server

   ollama serve

   Keep this terminal running.

4. Configure GitDeep

   • Settings → AI Provider → Local Ollama
   • Endpoint: http://localhost:11434
   • Model: qwen2.5:7b
   • Prompt Size: Full (if using 7B+ model)

Model Recommendations

Scenario	Model	RAM	Prompt Size	Employer Mode?
Best Overall	Gemini 2.5 Flash (cloud)	N/A	Full	✅ Yes
Premium Cloud	GPT-4o, Claude Sonnet 4	N/A	Full	✅ Yes
Local Strong	Llama 3.1 8B, Gemma 2 9B	16GB	Full	✅ Yes
Local Mid	Qwen 2.5 7B, Mistral 7B	8GB	Full	✅ Yes
Local Weak	Phi-3, Llama 3.2 3B	4GB	Small	❌ No
Too Small	Phi-1, TinyLlama	2GB	Small	❌ No

Important: Small models (< 7B parameters) produce inaccurate results including wrong scores, missing red flags, and useless analysis. Use cloud models or 7B+ local models for reliable assessments.

Custom Prompt for Ultra-Small Models

If your local model is still too small even for the "Small" prompt size:

1. Clone the repository locally (do NOT use the deployed website)
2. Use an external LLM (ChatGPT, Claude, Gemini) to shorten the prompts in lib/ai.ts
3. Find the buildPrompt() and buildSmallPrompt() functions
4. Ask the LLM: "Condense this assessment prompt to 200 tokens while keeping core evaluation criteria"
5. Replace the prompt strings in lib/ai.ts with the shortened versions
6. Run npm run dev and test at http://localhost:3000
7. CRITICAL: Do NOT deploy this modified version — it's for local testing only

⚠️ Warning: Ultra-compressed prompts will produce significantly less accurate assessments. This is a last resort for experimentation only.

---

📖 How It Works

Assessment Logic

1. Data Collection (via GitHub API)

   • User profile (bio, location, company, social links)
   • Top 15 non-fork repositories
   • README contents (first 1500 chars per repo)
   • Language statistics (actual byte counts)
   • Merged pull requests in external repos
   • Account age and activity patterns

2. AI Analysis (1200-token instruction prompt)

   • 3-tier band scoring system (1-4 Weak, 5-7 Average, 8-10 Strong)
   • Hard signals: (a) merged PRs, (b) original repos with READMEs, (c) 6+ months consistency
   • Behavioral pattern detection
   • Buzzword extraction and tech stack verification
   • AI slop detection vs quality orchestration
   • Career trajectory calculation
   • Per-repo independent scoring

3. Output Generation

   • Hirability score (1.0-10.0, never exactly 7.0)
   • SWOT analysis (all 4 fields always populated)
   • Detailed markdown report with structured sections
   • Radar charts for strengths/weaknesses
   • Timeline visualization
   • Role recommendations

Score Stability

The hirability score uses a fixed 3-tier band system to ensure consistency:

• Maximum deviation between assessments of the same profile: ±1.0
• Temperature locked to 0 for deterministic output
• Band determined by presence/absence of 3 hard signals
• Fine-tuning within band based on quality factors

---

⚠️ Known Limitations & Drawbacks

Current Issues

1. Small Model Performance

   • Models under 7B parameters (Phi, Llama 3.2 1B/3B, TinyLlama) produce unreliable results
   • Small prompt mode (400 tokens) sacrifices depth and accuracy
   • Status: Actively working on optimizing small prompts for employer mode

2. API Rate Limits

   • GitHub API: 60 requests/hour without token, 5000 with token
   • Search API (for PRs): 10 requests/min without token, 30 with token
   • Analyzing multiple candidates quickly can hit limits

3. PR Data Limitations

   • Only fetches top 15 merged PRs (GitHub Search API limit)
   • Cannot access PR diff details without individual API calls (would hit rate limits)
   • Lines added/deleted currently set to 0

4. README Truncation

   • Only first 1500 characters of each README analyzed
   • Large documentation may be incompletely assessed

5. Language Detection

   • Relies on GitHub's language detection (can be inaccurate for mixed projects)
   • Generated/vendor code may skew language percentages

6. Private Repo Analysis

   • Requires GitHub PAT token with appropriate scopes
   • Token must be manually entered (no OAuth flow)

7. Session Storage Limits

   • Assessed candidates stored in browser session storage (~5-10MB limit)
   • Clearing browser data loses comparison history
   • Private/incognito mode may block session storage

8. AI Hallucinations

   • AI may occasionally infer details not present in data
   • Behavioral analysis based on limited text samples (READMEs, bio)
   • Confidence scores are AI estimates, not measured metrics

9. No Real-Time Updates

   • Assessment is a snapshot at analysis time
   • Must re-run to see updated profile data

10. Context Window Constraints

    • Full prompt + profile data can exceed 8K tokens for prolific developers
    • May require switching to models with larger context (32K+)

Planned Improvements

• Optimize small prompt for employer mode accuracy
• Add caching layer for GitHub API responses
• Implement progressive README loading
• Add PR diff analysis (with rate limit handling)
• Export assessment reports as PDF/Markdown
• Add assessment history persistence (localStorage)
• Implement OAuth flow for GitHub authentication
• Add support for GitLab and Bitbucket profiles

---

🤝 Contributing

Help is appreciated! GitDeep is actively being improved, and contributions are welcome.

How to Contribute

1. Report Issues

   • Found a bug? Open an issue
   • Include: Steps to reproduce, expected vs actual behavior, browser/OS, AI provider used

2. Request Features

   • Have an idea? Open a feature request
   • Describe the use case and expected behavior

3. Submit Pull Requests

   • Fork the repository
   • Create a feature branch (git checkout -b feature/amazing-feature)
   • Commit changes (git commit -m 'Add amazing feature')
   • Push to branch (git push origin feature/amazing-feature)
   • Open a Pull Request

Current Focus Areas

• Small Model Optimization: Making 3B-7B models produce employer-grade assessments
• Prompt Engineering: Reducing token count while maintaining analysis depth
• Rate Limit Handling: Smarter API call batching and caching
• UI/UX Improvements: Better mobile responsiveness, loading states

Development Guidelines

• Follow existing code style (ESLint config provided)
• Test with multiple AI providers before submitting
• Update README if adding new features
• Keep client-side architecture (no server-side data storage)

---

🏗️ Architecture

Tech Stack

• Framework: Next.js 15 (App Router)
• Language: TypeScript 5.9
• Styling: Tailwind CSS 4.1
• Charts: Recharts 3.8
• GitHub API: Octokit 5.0
• AI SDKs: Google GenAI, native fetch for others
• Markdown: react-markdown 10.1

Project Structure

GitDeep/
├── app/
│   ├── page.tsx              # Home page (username input)
│   ├── assessment/page.tsx   # Assessment results page
│   ├── help/page.tsx         # Help & documentation
│   ├── layout.tsx            # Root layout with footer
│   ├── globals.css           # Global styles
│   └── api/ai/route.ts       # Server-side AI proxy (Ollama, OpenAI-compatible)
├── components/
│   └── SettingsModal.tsx     # Settings configuration UI
├── lib/
│   ├── ai.ts                 # AI provider logic & prompt engineering
│   ├── github.ts             # GitHub API data fetching
│   ├── store.tsx             # React Context for settings
│   ├── types.ts              # TypeScript interfaces
│   └── utils.ts              # Utility functions
├── public/
│   └── icon.svg              # Favicon
└── package.json

Data Flow

1. User enters GitHub username → app/page.tsx
2. Navigate to /assessment?user=username&mode=employer
3. fetchGitHubProfile() calls GitHub API → lib/github.ts
4. generateAssessment() sends data to AI → lib/ai.ts
5. AI response parsed and normalized → AssessmentResult
6. Results rendered with charts and markdown → app/assessment/page.tsx
7. Session storage saves candidate for comparison

---

📄 License

This project is licensed under the GNU General Public License v3.0 - see the LICENSE file for details. Any distributed modifications must remain open-source under the same terms.

---

👨💻 Developer

Yuvraj Sarathe

• 🐙 GitHub: @Yuvraj-Sarathe
• 🌐 Portfolio: yuvraj-sarathe.github.io/Portfolio
• 💼 LinkedIn: yuvraj-sarathe
• 🧩 LeetCode: Yuvraj_Sarathe

---

🙏 Acknowledgments

• GitHub API for comprehensive developer data
• Google Gemini for reliable free-tier AI access
• Ollama community for local model support
• Next.js team for excellent developer experience

---

📞 Support

If you find GitDeep useful, please ⭐ star the repository!

For questions or issues:

• Open an issue on GitHub Issues
• Check the Help Page for detailed documentation
• Try the live app at gitdeep.vercel.app

---

Built with ❤️ by developers, for developers

Try GitDeep Now | Report Bug | Request Feature