chiemekakalu/talentfilterdeployment

Startup Team Advisor - QLora finetuned Phi model

This model is a finetuned version of Microsoft's Phi-2 model, optimized for analyzing startup candidates, evaluating founding teams, and generating optimal team compositions.

Model Details

• Base Model: Microsoft Phi-2 (2.7B parameters)
• Finetuning Method: QLora (4-bit quantization with Low-Rank Adaptation)
• Domain: Startup team formation and analysis
• Use Cases:
  • Candidate skill assessment
  • Team composition analysis
  • Optimal founding team generation
  • Skill gap identification

Usage

This model is deployed as a Hugging Face Inference Endpoint with a custom handler that provides three main operations:

1. Candidate Analysis

{
  "inputs": {
    "operation": "analyze_candidate",
    "candidate": {
      "name": "Jane Doe",
      "skills": ["JavaScript", "React", "Node.js"],
      "experience": [
        {"title": "Senior Developer", "company": "Tech Company", "years": 3}
      ],
      "education": [
        {"institution": "Stanford", "degree": "MS Computer Science"}
      ]
    }
  }
}

2. Team Analysis

{
  "inputs": {
    "operation": "analyze_team",
    "team": [
      {
        "name": "Jane Doe",
        "skills": ["JavaScript", "React", "Node.js"]
      },
      {
        "name": "John Smith",
        "skills": ["Python", "Machine Learning", "Data Science"]
      }
    ],
    "include_startup_comparison": true
  }
}

3. Team Generation

{
  "inputs": {
    "operation": "generate_team",
    "candidates": [
      /* Array of candidate objects */
    ],
    "requirements": "Create a balanced team for a SaaS startup",
    "team_size": 5
  }
}

Response Format

All operations return responses with the same structure:

{
  "team_analysis": "Detailed text analysis...",
  "model_info": {
    "device": "cuda",
    "model_type": "phi-2-qlora-finetuned"
  }
}

The specific analysis field name will vary based on the operation (team_analysis, candidate_analysis, etc.).

Limitations

• This model works best with detailed candidate profiles
• Processing time increases with the number of candidates
• The model has a fallback mode if quantized loading fails
• Maximum context length is limited to approximately 2048 tokens

Implementation Details

The model uses:

• 4-bit quantization for efficient inference
• Phi-2 as the base model for high performance with fewer parameters
• Custom prompt templates optimized for team analysis
• Fallback mechanisms for graceful degradation