H-001 | AdPulse: Automated Insight Engine

Tagline: An automated reporting pipeline that converts raw ad campaign logs into executive-ready PDF briefings with AI-generated narratives in under 30 seconds.

---

1. The Problem (Real World Scenario)

Context: In the AdTech industry, reporting is a massive bottleneck. Account Managers currently spend 4-6 hours every week downloading CSVs from different DSPs, manually aggregating data in Excel, and taking screenshots to paste into PowerPoint decks.

The Pain Point:

1. Latency: Data is often days old by the time the client sees it.
2. Error Rate: Manual copy-pasting leads to human error in critical metrics like Spend and CPA.
3. Wasted Talent: Highly paid strategists are doing low-value data entry work instead of optimizing campaigns.

My Solution: I built AdPulse, a streamlined insight engine. The user uploads a raw data file, and the system instantly standardizes the data, calculates the math deterministically, and uses Generative AI to write a strategic summary. The output is a finalized PDF ready to be emailed to a CMO.

---

2. Expected End Result

For the User:

• Input: Upload a raw CSV file (Campaign Performance Report).
• Action: Click "Generate Executive Report".
• Output: A downloadable, professional PDF containing:
  • Hard KPIs (Total Spend, Conversions, CPA) calculated with 100% precision.
  • An AI-written executive summary explaining why a specific campaign is winning or losing.

---

3. Technical Approach

I designed this system to prioritize accuracy and speed, moving away from "chatbots" to a functional business tool.

System Architecture:

1. Ingestion Layer (Streamlit):

   • Acts as the frontend interface for secure file handling.
   • Provides immediate visual feedback and data preview.

2. Data Engine (Pandas):

   • Decision: I chose Pandas over pure AI processing for the calculations.
   • Reasoning: LLMs are probabilistic and often fail at basic arithmetic (hallucinations). I strictly used Pandas for all aggregations (Sum, Avg, CPA calculation) to ensure financial accuracy.

3. Intelligence Layer (Google Gemini):

   • We pass the aggregated metadata (not the raw rows) to Google Gemini.
   • Prompt Engineering: Used a "Role-Playing" prompt to force the model to act as a Senior Analyst, focusing on actionable insights rather than generic observations.

4. Reporting Engine (FPDF):

   • Programmatically renders the final PDF. This ensures a consistent corporate layout that is difficult to achieve with simple HTML-to-PDF converters.

---

4. Tech Stack

• Language: Python 3.10+
• Frontend: Streamlit
• Data Processing: Pandas (NumPy backend)
• AI Model: Google Gemini API (via LangChain)
• Document Generation: FPDF
• Environment Management: Python Dotenv

---

5. How to Run

1. Clone Repository

   git clone https://github.com/manojkp08/GroundTruthAIHackathon2025_Manoj.git

2. Install Dependencies

   pip install -r requirements.txt

3. Configure API Key Create a .env file in the root directory:

   GEMINI_API_KEY="your_google_api_key_here"

4. Run the Application

   streamlit run app.py

5. Test with Sample Data

   • Use the custom_csv/campaign_data.csv provided in the repo to test the pipeline.