AgentFlow

AI engine that transform raw, messy dataset into executive business intelligence in second. It automatically cleans data, detect Z-score anomalies, generate C-Suite PDF reports with zero manual effort

Comment

In today's data-driven world, modern enterprise organizations drown in massive, unorganized datasets (.xlsx, .csv, .json, .parquet). Translating raw, messy tabular data into C-Suite executive intelligence currently requires hours or days of manual data cleaning, statistical profiling, anomaly detection, and report design.

We envisioned AgentFlow — an autonomous multi-agent AI data analytics engine powered by Google Gemini 2.5 Flash and built on the Google Agent Framework. AgentFlow automates the entire analytical lifecycle, ingesting raw datasets and producing fully-validated executive PDF reports, real-time Chart.js dashboards, and interactive AI Q&A assistants in seconds.

AgentFlow transforms raw enterprise data into actionable business intelligence through an autonomous 10-Agent Directed Acyclic Graph (DAG):

  • 🤖 Autonomous 10-Agent DAG Orchestration: Ten specialized AI agents (Planner, Ingestion, Profiler, Cleaning, Anomaly, KPI, Trend, Validator, Insight, Report) collaborate in parallel without human intervention.
  • 📁 Universal Multi-Format Ingestion: Native parser supporting .xlsx, .csv, .json, .parquet, and .tsv files up to 500MB with automated UTF-8 BOM, latin-1, and NaN sanitization.
  • 📐 $3\sigma$ Z-Score Anomaly Engine: Identifies critical statistical outliers across numeric columns using the standard Z-score model: $$Z = \frac{x - \mu}{\sigma}$$ where $x$ is the data value, $\mu$ is the population mean, and $\sigma$ is the standard deviation.
  • 💬 "Ask AgentFlow" AI Assistant: Interactive natural language Q&A interface powered by Google Gemini 2.5 Flash via the google-genai SDK.
  • 📄 1-Click Executive PDF Generation: Programmatic ReportLab PDF engine compiling data quality scores, KPI summary matrices, anomaly highlights, and strategic recommendations.
  • 🌐 3-Language i18n Engine: Live real-time translation support for English (Default), Hindi, and Spanish.

- AI Core & Agent Framework: Integrated Google Gemini 2.5 Flash (gemini-2.5-flash) using the official google-genai Python SDK (GeminiProvider), structured under Google Agent Framework specifications.

  • Backend Infrastructure: Django 5.2, Django REST Framework, Django Channels (WebSockets for real-time agent execution tracking), Celery/Redis for asynchronous worker tasks, and Python 3.14.
  • Data Engineering Engine: Pandas, NumPy, PyArrow, OpenPyXL, and ReportLab.
  • Google Cloud Platform Services:
    • Google Cloud Run: Containerized serverless deployment using Docker, cloudbuild.yaml, and automated shell scripts (deploy_gcp.sh / deploy_gcp.ps1).
    • Google Cloud Storage (GCS): Multi-bucket cloud persistence (gs://agentflow-datasets and gs://agentflow-results).
    • Google Secret Manager: Secure API key and secret retrieval (config/secrets.py).
  • Frontend & UI: HTML5, Google Material 3 CSS, Vanilla JS, and Chart.js v4.x.

1. DAG Synchronization & Parallel Latency: Guaranteeing parallel execution of independent analysis agents (AnomalyAgent, KPIAgent, TrendAgent) while maintaining strict sequential dependencies for downstream agents (ValidatorAgent, InsightAgent, ReportAgent).

  1. Multi-Format Data Sanitization: Handling edge-case encoding errors (UTF-8 BOM vs latin-1), missing values, and converting NumPy/Pandas data types and NaN/Inf values into clean JSON structures for Gemini API prompts.
  2. Zero-Dependency Local & Cloud Hybrid Execution: Designing an architecture that runs 100% locally out-of-the-box with zero mandatory setup, while seamlessly scaling to Google Cloud Run and Google Gemini 2.5 Flash when cloud credentials are provided.

- 100% Test Suite Coverage: Passed 13 out of 13 unit tests and end-to-end integration smoke tests (pytest).

  • Seamless Gemini 2.5 Integration: Built native support for gemini-2.5-flash using Google's newest google-genai Python SDK.
  • Sub-5-Second Execution: Achieved end-to-end 10-agent workflow execution, anomaly detection, AI insight synthesis, and PDF generation in under 5 seconds for 100,000-row datasets.
  • Production-Ready Deployment: Created 1-command deployment scripts for Google Cloud Run (deploy_gcp.ps1 / deploy_gcp.sh).

- Deep practical experience building autonomous multi-agent DAG architectures using the Google Agent Framework and Google Gemini 2.5 Flash.

  • Best practices for structured JSON output formatting (response_mime_type: "application/json") with google-genai SDK for analytical pipelines.
  • Strategies for containerizing complex multi-threaded Python applications (Django + Daphne + WebSockets + Celery) for serverless deployment on Google Cloud Run.

- Google BigQuery & Spanner Connectors: Direct native integration for querying massive enterprise data warehouses directly from AgentFlow.

  • Multi-Modal Gemini Vision Integration: Analyzing uploaded charts, infographics, and PDF slide decks alongside raw tabular datasets.
  • Automated Enterprise Webhooks: Pushing generated executive PDF reports to Slack, Microsoft Teams, and email automatically upon workflow completion.

Built With

Share this project:

Updates

posted an update —

Excited to announce the official launch of AgentFlow for the Google AI Hackathon! AgentFlow is an enterprise-grade autonomous multi-agent AI data analytics engine that transforms raw, messy tabular datasets (.xlsx, .csv, .json, .parquet) into executive-ready business intelligence in seconds. Key Milestones:

  • Integrated Google Gemini 2.5 Flash via the official google-genai SDK
  • Built 10-Agent Autonomous DAG Pipeline using Google Agent Framework standards
  • Production-ready deployment for Google Cloud Run, Google Cloud Storage (gs://), & Google Secret Manager
  • Real-time $3\sigma$ Z-score anomaly detection ($Z = \frac{x - \mu}{\sigma}$)
  • 1-Click C-Suite Executive PDF report generation GitHub Repository: https://github.com/chandan326/agentflow

Log in or sign up for Devpost to join the conversation.

Submission history