Deep Dive: Building an AI-Driven Balance Sheet Analyzer with Streamlit & Gemini

Financial analysis is often gate-kept by complex spreadsheets and static PDF reports. In this post, we’ll break down the technical architecture of a “Premium Balance Sheet Analyzer” that uses Generative AI to bridge the gap between static financial statements and real-time market data.

---

1. Unified Data Ingestion Architecture

One of the most challenging aspects of financial automation is handling varied input sources. Our application implements a Dual-Stream Ingestion Path for both local files and remote URLs.

Handling Remote Streams with io.BytesIO

To avoid bloating server storage, we process remote URLs in-memory using the requests library and io.BytesIO.

import requests
import io

response = requests.get(file_url)
response.raise_for_status()
source_file = io.BytesIO(response.content)
source_file.name = "downloaded_report.pdf" # Metadata normalization

PDF vs. Excel Parsing Strategies

• PDF Extraction: We use pdfplumber instead of basic PyPDF2 because it handles tabular data with higher fidelity, preserving the vertical alignment of currency columns.
• Excel Extraction: We use pd.read_excel(file, sheet_name=None) to iterate through every sheet in the workbook, converting numerical tables into a sanitized text representation that the LLM can ingest easily.

---

2. Intelligence Layer: Gemini 2.5 Flash

We selected Gemini 2.5 Flash for its massive context window (over 1M tokens) and multimodal reasoning capabilities.

Complex Multi-Stage Prompting

The core logic resides in a two-stage inference process:

Stage 1: Extraction & Ticker Identification

The AI is instructed to identify the “Probable Ticker Symbol” for the Indian stock exchange. We use a standardized output format: TICKER: <SYMBOL>.NS.

prompt = """
Analyze the data for:
1. Key Ratios (Current Ratio, D/E, ROE)
2. Red Flags
...
Identify the NSE/BSE ticker symbol. Output as 'TICKER: SYMBOL.NS'
"""

Stage 2: Recursive Valuation

We take the output of Stage 1, fetch real-time data from yfinance, and then pass both the original balance sheet and the live market data back to Gemini to determine if the stock price is “justified.”

---

3. Real-Time Market Integration with yfinance

Once the ticker is identified (e.g., TCS.NS), we hit the Yahoo Finance API.

stock = yf.Ticker("RELIANCE.NS")
info = stock.info
curr_price = info.get('currentPrice')
# Fetching multiples like P/E and Market Cap
market_cap = info.get('marketCap') / 1e7 # Converting to Cr (Standard Indian Finance)

We use this data to populate Plotly Graph Objects, creating a dynamic 30-day price action chart that provides context to the balance sheet health.

---

4. State Management and Conversations

Streamlit’s st.session_state is used to maintain the chat history and the analysis results, preventing redundant API calls (and costs).

Chat History Management

Gemini’s start_chat method is perfect for keeping financial context. We store the history object in the session state.

# Initializing Chat with context
if "chat_history" not in st.session_state:
    st.session_state.chat_history = []

# Appending with history awareness
chat = model.start_chat(history=st.session_state.chat_history)
response = chat.send_message(user_query)
st.session_state.chat_history = chat.history

---

5. Premium UI Engineering

To achieve a “Bloomberg Terminal” aesthetic, we bypassed standard Streamlit layouts with custom CSS and glassmorphism.

Glassmorphism Card System

.card {
    background: rgba(30, 41, 59, 0.7);
    backdrop-filter: blur(12px);
    border: 1px solid rgba(255, 255, 255, 0.1);
    border-radius: 16px;
    padding: 24px;
}

Multi-Tab Navigation

We use st.tabs to separate the Qualitative Report, Quantitative Valuation, and the Interactive AI Chat. This keeps the user focused on one dimension of the data at a time while preserving the overall context.

---

🛡️ Best Practices & Security

• API Key Management: Supports both .env variables and runtime user input (stored safely in session).
• Graceful Failures: Wrapped extraction logic in try-except blocks to handle malformed PDF tables.
• Resource Cleanup: Used io.BytesIO to ensure no temporary files are left on the execution server.

This implementation demonstrates the power of combining modern LLMs with real-time financial APIs to build institutional-grade tools on top of lightweight frameworks like Streamlit.

Github link – https://github.com/sethlahaul/streamlit-balance-sheet-analyzer