Bengaluru AI Job Radar

Bengaluru AI Job Radar is a Python FastMCP server that helps an AI agent search for AI internship and early-career AI roles in Bengaluru, save them to a local JSON job tracker, and render a Prefab dashboard UI.

Built with FastMCP 3.4.x, Prefab UI 0.20.x, Tavily for internet search, and JSON for local persistence.

Assignment Mapping

Architecture

User prompt
  → Agent calls search_ai_jobs (Tavily internet search)
  → Agent saves results via job_tracker_db (JSON CRUD)
  → Agent reads records via job_tracker_db (JSON CRUD)
  → Agent calls render_job_dashboard (Prefab UI)
  → Prefab dashboard appears in MCP host

MCP Tools

Setup (Windows)

cd C:\Cursor\EAGv3\S4
cd bengaluru-ai-job-radar

python -m venv .venv
.venv\Scripts\activate

pip install -e ".[dev]"

copy .env.example .env
# Edit .env and add your TAVILY_API_KEY

Setup (macOS / Linux)

cd /path/to/bengaluru-ai-job-radar

python3 -m venv .venv
source .venv/bin/activate

pip install -e ".[dev]"

cp .env.example .env
# Edit .env and add your TAVILY_API_KEY

Environment Variables

Create a .env file (or set system environment variables):

TAVILY_API_KEY=tvly-xxxxxxxxxxxxxxxxx
DATABASE_PATH=data/bengaluru_ai_job_radar.db

• `TAVILY_API_KEY` (required): Get one free at tavily.com.
• `DATABASE_PATH` (optional): Defaults to data/bengaluru_ai_job_radar.db relative to the project root.

Running the MCP Server

Direct Python execution

python -m bengaluru_ai_job_radar.server

Using FastMCP CLI

fastmcp run src/bengaluru_ai_job_radar/server.py

App preview (if supported)

fastmcp dev src/bengaluru_ai_job_radar/server.py

Connecting to an MCP Host

Add this to your MCP host configuration (Claude Desktop, Cursor, VS Code, etc.):

{
  "mcpServers": {
    "bengaluru-ai-job-radar": {
      "command": "python",
      "args": ["-m", "bengaluru_ai_job_radar.server"],
      "cwd": "C:\\Cursor\\EAGv3\\S4\\bengaluru-ai-job-radar",
      "env": {
        "TAVILY_API_KEY": "your_key_here",
        "DATABASE_PATH": "data/bengaluru_ai_job_radar.db"
      }
    }
  }
}

Note: Exact MCP host configuration may differ depending on Claude Desktop, Cursor, VS Code, ChatGPT MCP Apps, or another host.

Demo Prompt

Copy this prompt into your MCP-connected AI agent to exercise all 3 tools:

Use the Bengaluru AI Job Radar MCP server to complete this full workflow.

>

First, use the MCP internet search tool search_ai_jobs to find companies currently hiring AI Interns, ML Interns, GenAI Interns, LLM Engineer Interns, AI Implementation Engineer Interns, or Junior AI Engineers in Bengaluru.

>

Prioritize roles involving Python, LLMs, RAG, AI agents, embeddings, prompt engineering, fine-tuning, model training, NLP, or AI backend development.

>

Save at least 5 relevant job leads to the local JSON job tracker using the MCP CRUD tool job_tracker_db.

>

Then read the saved records back using job_tracker_db with the list_jobs operation.

>

Finally, render the saved results using the FastMCP Prefab UI tool render_job_dashboard.

>

Do not answer from memory. Do not skip any step. You must call all 3 tools: 1. search_ai_jobs 2. job_tracker_db 3. render_job_dashboard

The full prompt is also available in demo_prompt.md.

Example Workflow

1. Search for roles

The agent calls search_ai_jobs with:

{
  "role_query": "AI Intern",
  "location": "Bengaluru",
  "max_results": 10
}

2. Save job leads

The agent calls job_tracker_db for each result:

{
  "operation": "create_job",
  "payload": {
    "company": "Sarvam AI",
    "role_title": "AI Intern",
    "role_type": "internship",
    "location": "Bengaluru",
    "skills": ["Python", "LLM", "RAG"],
    "fit_score": 85,
    "source_platform": "LinkedIn",
    "source_url": "https://linkedin.com/jobs/view/123"
  }
}

3. List saved jobs

{
  "operation": "list_jobs",
  "payload": {"min_fit_score": 50}
}

4. Update a role status

{
  "operation": "update_status",
  "payload": {
    "job_id": "...",
    "new_status": "applied",
    "event_note": "Applied via company careers page"
  }
}

5. Add a note

{
  "operation": "add_note",
  "payload": {
    "job_id": "...",
    "note": "Reach out to founder on LinkedIn"
  }
}

6. Render dashboard

The agent calls render_job_dashboard → a Prefab UI dashboard appears with summary cards, a job table, recent activity, and skill frequency.

Running Tests

pytest tests/ -v

Tests cover:

• Fit score calculator — scoring rubric, caps, keyword boosts
• Normalization — company names, slug IDs, role type inference, skill extraction
• JSON store — full CRUD lifecycle, upsert, deduplication, dashboard aggregation

Project Structure

bengaluru-ai-job-radar/
  README.md
  pyproject.toml
  .env.example
  .gitignore
  demo_prompt.md

  src/
    bengaluru_ai_job_radar/
      __init__.py
      server.py          # FastMCP server entrypoint
      config.py           # Environment variable loading
      schemas.py          # Pydantic validation models

      tools/
        __init__.py
        search.py          # search_ai_jobs MCP tool
        database.py        # job_tracker_db MCP tool
        dashboard.py       # render_job_dashboard MCP tool (Prefab UI)

      services/
        __init__.py
        tavily_service.py  # Tavily API integration
        fit_score.py       # Deterministic fit-score calculator
        normalization.py   # Company/role normalization, skill extraction

      storage/
        __init__.py
        json_store.py      # JSON CRUD store (JobRadarStore)

  data/
    .gitkeep              # JSON DB created here at runtime

  tests/
    test_fit_score.py
    test_normalization.py
    test_json_store.py

Known Limitations

• Tavily search result quality depends on public indexing of job boards.
• Some job platforms may block direct scraping, so the tool relies on Tavily snippets and source links.
• Compensation data may be unavailable for many listings — shown as "unknown".
• Prefab UI APIs are actively evolving; dependency pinning to prefab-ui>=0.20.0,<1.0.0 is used.
• Dashboard interactivity depends on the MCP host's support for FastMCP Apps / Prefab rendering.
• Company name extraction from search results is best-effort; some may show as "unknown".
• The fit-score algorithm is deterministic and rule-based — it does not use ML or LLM reasoning.

Dependencies

License

MIT