# findata-mcp

**Financial Data Quality & AI Inference Evaluation MCP Server**

A Model Context Protocol (MCP) server that exposes six production-grade tools for AI agents working with financial datasets — covering data quality auditing, bias detection, model inference evaluation, outlier scoring, A/B testing, and KPI reporting.

Built to mirror the core responsibilities of Citi's Data Services & AI platform.

---

## Tools

| Tool | Description |
|---|---|
| `audit_data_quality` | Audits completeness, consistency, and machine-readability of financial records. Returns a quality score and remediation actions. |
| `detect_bias` | Detects demographic/categorical bias by comparing approval rates or amounts across cohort groups. Returns disparity ratios and a bias risk label (LOW / MEDIUM / HIGH). |
| `evaluate_model_inference` | Computes precision, recall, F1, AUC, and a PASS/FAIL verdict against configurable enterprise thresholds. |
| `score_outliers` | Flags anomalous records using Z-score and IQR methods across numeric fields. |
| `run_ab_comparison` | Compares two model variants on the same dataset and recommends a winner based on F1. |
| `generate_kpi_report` | Generates a structured KPI report — latency (avg/p95), throughput (rps), error rate, and SLA adherence — formatted for senior stakeholder delivery. |

---

## Installation

```bash
git clone https://github.com/srikarmanikonda/findata-mcp.git
cd findata-mcp
npm install
```

## Run

```bash
node src/index.js
```

The server communicates over **stdio** using the MCP protocol — connect it to any MCP-compatible client (Claude Desktop, ADK agent, etc.).

## Test

```bash
npm test
```

All 13 tests pass across all 6 tools.

---

## Claude Desktop / MCP Client Config

Add to your `claude_desktop_config.json`:

```json
{
  "mcpServers": {
    "findata-mcp": {
      "command": "node",
      "args": ["/path/to/findata-mcp/src/index.js"]
    }
  }
}
```

---

## Example Usage (via MCP client)

**Audit data quality:**
```json
{
  "tool": "audit_data_quality",
  "records": [
    { "id": "R001", "loan_amount": 15000, "region": "Northeast", "approved": 1 },
    { "id": "R002", "loan_amount": null,  "region": "Southeast", "approved": 0 }
  ],
  "required_fields": ["id", "loan_amount", "region", "approved"],
  "numeric_fields": ["loan_amount"]
}
```

**Detect bias:**
```json
{
  "tool": "detect_bias",
  "records": [...],
  "group_field": "region",
  "outcome_field": "approved",
  "outcome_type": "binary"
}
```

**Evaluate model inference:**
```json
{
  "tool": "evaluate_model_inference",
  "predictions": [
    { "id": "R001", "predicted": 0.91, "actual": 1 },
    { "id": "R002", "predicted": 0.22, "actual": 0 }
  ],
  "threshold": 0.5,
  "min_precision": 0.75,
  "min_recall": 0.70
}
```

---

## Stack

- **Runtime:** Node.js (ESM)
- **MCP SDK:** `@modelcontextprotocol/sdk`
- **Validation:** `zod`
- **Transport:** stdio (MCP standard)

---

## Author

Srikar Manikonda — [srikarmanikonda9@gmail.com](mailto:srikarmanikonda9@gmail.com)