Data Pipeline Agent#

Build, run, and monitor ETL (Extract → Transform → Load) pipelines for business data. Specializes in financial data flows, API integrations, and warehouse loading patterns for accounting and operations teams.

When to Use#

• Extracting data from APIs (QBO, Stripe, Salesforce, bank feeds, etc.)
• Cleaning and normalizing messy spreadsheets or CSV exports
• Merging data from multiple sources into one canonical dataset
• Loading transformed data into databases, data warehouses, or Google Sheets
• Scheduling recurring data syncs (daily GL pulls, weekly AR aging refresh, etc.)
• Auditing data quality — detecting nulls, duplicates, type mismatches

When NOT to Use#

• Real-time streaming — use Kafka, Kinesis, or Pub/Sub for sub-second latency
• Interactive dashboards — this agent outputs data; visualization belongs in BI tools
• Raw SQL query optimization — use DBA tooling for query plans and indexes
• One-off manual exports — if it happens once, just download the CSV
• Transactional writes to client systems — read-only extraction only unless Irfan approves write access

---

Pipeline Patterns#

Pattern 1: API Extract → Clean → CSV#

python

# Extract from REST API, clean, output CSV
import requests, pandas as pd, json
from datetime import datetime, timedelta

def extract(api_url, headers, params=None):
    """Pull paginated JSON from any REST endpoint."""
    results = []
    while api_url:
        r = requests.get(api_url, headers=headers, params=params)
        r.raise_for_status()
        data = r.json()
        results.extend(data.get("data", data if isinstance(data, list) else [data]))
        api_url = data.get("next_page_url")  # pagination
        params = None  # only pass params on first call
    return results

def clean(records, rename_map=None, drop_nulls_on=None, date_cols=None):
    """Normalize, rename, parse dates, drop nulls."""
    df = pd.DataFrame(records)
    if rename_map:
        df = df.rename(columns=rename_map)
    if date_cols:
        for col in date_cols:
            df[col] = pd.to_datetime(df[col], errors="coerce")
    if drop_nulls_on:
        df = df.dropna(subset=drop_nulls_on)
    df = df.drop_duplicates()
    return df

def load_csv(df, output_path):
    df.to_csv(output_path, index=False)
    print(f"✅ Saved {len(df)} rows → {output_path}")

# Example: QBO Invoice Extract
HEADERS = {"Authorization": "Bearer <TOKEN>", "Accept": "application/json"}
records = extract("https://quickbooks.api.intuit.com/v3/company/<REALM>/query?query=SELECT * FROM Invoice", HEADERS)
df = clean(records, rename_map={"TxnDate": "invoice_date", "TotalAmt": "amount"}, date_cols=["invoice_date"])
load_csv(df, f"data/invoices_{datetime.today().date()}.csv")

Pattern 2: Multi-Source Merge#

python

import pandas as pd

def merge_gl_with_bank(gl_path, bank_path, match_on="amount", date_tolerance_days=3):
    """
    Match GL entries to bank transactions.
    Flags unmatched rows for manual review.
    """
    gl = pd.read_csv(gl_path, parse_dates=["date"])
    bank = pd.read_csv(bank_path, parse_dates=["date"])

    # Merge on amount + date proximity
    merged = pd.merge_asof(
        gl.sort_values("date"),
        bank.sort_values("date"),
        on="date",
        by=match_on,
        tolerance=pd.Timedelta(days=date_tolerance_days),
        direction="nearest",
        suffixes=("_gl", "_bank")
    )

    unmatched_gl = gl[~gl.index.isin(merged.dropna(subset=["date_bank"]).index)]
    unmatched_bank = bank[~bank.index.isin(merged.dropna(subset=["date_gl"]).index)]

    print(f"✅ Matched: {len(merged.dropna())} | ⚠️ Unmatched GL: {len(unmatched_gl)} | Bank: {len(unmatched_bank)}")
    return merged, unmatched_gl, unmatched_bank

Pattern 3: Data Quality Audit#

python

import pandas as pd

def audit_dataset(df, required_cols=None, expected_types=None):
    """
    Run data quality checks. Returns a report dict.
    """
    report = {
        "row_count": len(df),
        "duplicate_rows": int(df.duplicated().sum()),
        "null_summary": df.isnull().sum().to_dict(),
        "issues": []
    }

    if required_cols:
        missing = [c for c in required_cols if c not in df.columns]
        if missing:
            report["issues"].append(f"Missing required columns: {missing}")

    if expected_types:
        for col, dtype in expected_types.items():
            if col in df.columns and not pd.api.types.is_dtype_equal(df[col].dtype, dtype):
                report["issues"].append(f"{col}: expected {dtype}, got {df[col].dtype}")

    # Flag columns with >20% nulls
    for col, nulls in report["null_summary"].items():
        pct = nulls / len(df) * 100
        if pct > 20:
            report["issues"].append(f"{col}: {pct:.1f}% null — review required")

    return report

# Usage
df = pd.read_csv("data/ar_aging.csv")
report = audit_dataset(
    df,
    required_cols=["customer_id", "invoice_date", "amount", "due_date"],
    expected_types={"amount": "float64", "customer_id": "object"}
)
print(report)

Pattern 4: Scheduled Cron Pipeline#

bash

#!/bin/bash
# daily-gl-sync.sh — run via cron or OpenClaw cron tool
# Extracts GL, cleans, loads to SQLite, notifies on error

set -euo pipefail
LOG="logs/gl-sync-$(date +%Y-%m-%d).log"
mkdir -p logs data

echo "[$(date)] Starting GL sync..." | tee -a "$LOG"

python3 pipelines/gl_extract.py >> "$LOG" 2>&1 && \
python3 pipelines/gl_clean.py >> "$LOG" 2>&1 && \
python3 pipelines/gl_load.py >> "$LOG" 2>&1 && \
echo "[$(date)] ✅ GL sync complete" | tee -a "$LOG" || \
echo "[$(date)] ❌ GL sync FAILED — check $LOG" | tee -a "$LOG"

Pattern 5: Load to SQLite / PostgreSQL#

python

import pandas as pd
import sqlite3

def load_to_sqlite(df, db_path, table_name, if_exists="replace"):
    """
    Load DataFrame to SQLite. Use if_exists='append' for incremental loads.
    """
    conn = sqlite3.connect(db_path)
    df.to_sql(table_name, conn, if_exists=if_exists, index=False)
    conn.close()
    print(f"✅ Loaded {len(df)} rows → {db_path}::{table_name}")

# PostgreSQL version (requires psycopg2 + sqlalchemy)
from sqlalchemy import create_engine

def load_to_postgres(df, conn_str, table_name, schema="public", if_exists="replace"):
    engine = create_engine(conn_str)
    df.to_sql(table_name, engine, schema=schema, if_exists=if_exists, index=False)
    print(f"✅ Loaded {len(df)} rows → {schema}.{table_name}")

---

Common Business Pipelines#

AR Aging Refresh Pipeline#

text

1. Extract: QBO Invoices API → raw JSON
2. Transform: Calculate days_outstanding, aging_bucket (0-30, 31-60, 61-90, 90+)
3. Enrich: Join with customer contact data
4. Load: Google Sheets "AR Aging" tab + SQLite archive
5. Alert: Flag invoices >60 days for follow-up queue

Bank Feed Reconciliation Pipeline#

text

1. Extract: Bank API (Plaid/CSV export) + QBO GL
2. Transform: Normalize dates, amounts, memo fields
3. Match: Fuzzy join on amount + date (±3 days tolerance)
4. Flag: Unmatched transactions → manual review CSV
5. Load: Reconciliation log → SQLite + email summary

Payroll → GL Mapping Pipeline#

text

1. Extract: Payroll system CSV export (Gusto, ADP, etc.)
2. Transform: Map payroll codes → GL account numbers
3. Validate: Totals match payroll register
4. Load: Journal entry template → QBO batch import format
5. Archive: Raw + transformed files in dated folder

---

Pipeline Design Checklist#

Before building any pipeline:

• Idempotency — Can the pipeline re-run without duplicating data?
• Error handling — What happens if the API is down? Partial load?
• Logging — Is every step logged with timestamps?
• Data quality — Are nulls, duplicates, and type mismatches caught?
• Reversibility — Can the load be rolled back if something goes wrong?
• Rate limits — Does the source API have call limits? Add retry logic.
• Secrets — Are API keys in env vars, not hardcoded?
• Schedule — How often does this run? Who monitors it?

---

Data Cleaning Quick Reference#

ProblemSolutionMixed date formatspd.to_datetime(col, infer_datetime_format=True)Currency strings ("$1,234.56")col.str.replace(r'[$,]', '', regex=True).astype(float)Duplicate rowsdf.drop_duplicates(subset=['id'])Null amountsdf['amount'].fillna(0) or df.dropna(subset=['amount'])Inconsistent casingdf['name'].str.strip().str.title()Leading/trailing spacesdf.apply(lambda x: x.str.strip() if x.dtype == 'object' else x)Outlier detectiondf[df['amount'].between(df['amount'].quantile(.01), df['amount'].quantile(.99))]

---

Scheduling with OpenClaw Cron#

text

# Daily GL sync at 6 AM CST
Schedule: cron "0 6 * * *" tz=America/Chicago
Payload: agentTurn — "Run the daily GL sync pipeline in ~/workspace/pipelines/"
Delivery: announce to Telegram on completion or failure

---

Dependencies#

Install Python data stack:

bash

pip install pandas requests sqlalchemy psycopg2-binary openpyxl xlrd
# For Google Sheets
pip install gspread gspread-dataframe google-auth
# For Plaid bank feeds
pip install plaid-python

---

File Organization#

text

workspace/
  pipelines/
    gl_extract.py
    gl_clean.py
    gl_load.py
    ar_aging.py
    bank_reconcile.py
  data/
    raw/          ← API responses, CSV imports (never edited)
    processed/    ← cleaned, transformed data
    archive/      ← date-stamped historical snapshots
  logs/
    pipeline-YYYY-MM-DD.log
  scripts/
    run-daily-pipelines.sh

---

Safety Rules#

1. Extract is always read-only. Never write to source systems during extraction.
2. Archive raw data before any transformation — keep the original.
3. Validate row counts before and after each transformation step.
4. Test on sample data (10-100 rows) before running full pipeline.
5. Client system writes require Irfan approval — QBO, bank APIs, payroll systems are extract-only by default.
6. Never hardcode credentials — use environment variables or 1Password CLI.