Oculon: Delhi Hotspots ML Explorer

Flexible User-Controlled Query Engine

Select what kind of data you need and how much data to return (any count or the complete dataset). Runs automatic independent related analysis across multi-year history.

Click 'Execute User-Controlled Query' to view records.

Dynamic On-The-Fly Model Execution

Run the spatial multigram model dynamically for any month using the complete historical event stream. Evaluate any custom count of units or all 210 candidate police stations.

Choose parameters and click 'Run Live Model Forecast'.

Dual Evaluation & Oracle Perplexity Framework

The Delhi Hotspots ML system is evaluated on two rigorous, independent benchmarks:

1. Preserved Baseline (80/20 Random Holdout): Historical holdout evaluation.
2. Strict Forward-Time (Expanding-Window Rolling Origin): Reserves the latest 20% of complete calendar months, strictly forecasting each test month using only past events.
3. Theoretical Oracle Floor: Evaluated at 125.82 perplexity across 27 holdout months.

1. Preserved 80/20 Holdout Baseline

2. Forward-Time Rolling-Origin Evaluation (27 Complete Test Months)

Connect AI Agents to Delhi Hotspots ML via MCP

This Hugging Face Space functions as a complete remote MCP Server that can run independent related queries and live model inference on the complete datasets.

1. Claude Desktop Configuration

Supports both Streamable HTTP (recommended, modern MCP specification) and legacy SSE:

Option A: Streamable HTTP via mcp-remote (Recommended)

{
  "mcpServers": {
    "oculon_delhi_hotspots": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://abhyudaymishr-oculon.hf.space/mcp",
        "--transport",
        "http-only"
      ]
    }
  }
}

Option B: Claude Remote Connector (Streamable HTTP) Under Claude Desktop Connectors -> Advanced -> Transport: Select Streamable HTTP, URL: https://abhyudaymishr-oculon.hf.space/mcp

Option C: Legacy SSE Transport https://abhyudaymishr-oculon.hf.space/sse

2. Available Universal MCP Tools

• fetch_dataset_records(dataset, data_type, month, limit, offset, filter_district, filter_station): Completely user-controlled fetching (any limit or 'all') for any data type.
• query_hotspots(dataset, month, query_text, top_k): Query complete datasets with user-chosen top_k.
• predict_hotspots(dataset, target_month, top_k): Run live model prediction on-the-fly for any month.
• query_dataset_analytics(dataset, query_type, entity_name, limit): Historical longitudinal aggregations.
• get_hotspot_map(map_type): Direct URLs to interactive Leaflet hotspot maps.
• get_model_metrics(dataset, evaluation_type): Dual-evaluation benchmarks and oracle perplexity floor.

3. Programmatic REST API

• GET /api/hotspots?dataset=missing_persons&top_k=50
• GET /api/fetch_records?dataset=stolen_vehicles&data_type=all_units&limit=all
• POST /api/query_hotspots (body: {"dataset": "unidentified_bodies", "top_k": "all"})
• POST /api/fetch_records (body: {"dataset": "missing_persons", "data_type": "monthly_timeline", "limit": "all"})
• POST /api/predict_hotspots (body: {"dataset": "missing_persons", "target_month": "2026-12", "top_k": 210})