DocuTrace: Verifiable AI Auditor

Precision Data Extraction with Pixel-Perfect Source Grounding

View Live Demo • Engineering Journey • Architecture

---

Executive Summary

DocuTrace is a specialized auditing tool designed for high-stakes domains (Legal, Finance, Compliance) where AI hallucinations are unacceptable.

Unlike standard RAG (Retrieval Augmented Generation) pipelines that summarize text, DocuTrace utilizes Google's LangExtract library to perform Source Grounding. Every extracted data point is cryptographically linked to its specific coordinates in the source PDF, generating an interactive HTML audit trail.

---

Engineering Journey: Challenges & Solutions

Building a verifiable extraction engine involves overcoming significant "dependency hell" and rate-limiting barriers.

1. The "Slim Image" Dependency Conflict

• The Challenge: We deployed using the standard python:3.10-slim Docker image to minimize boot time. However, the LangExtract library requires installation directly from GitHub source, which depends on git. The build failed with ExecutableNotFound: git.
• The Solution: Architected a custom multi-stage Dockerfile that injects system-level dependencies (apt-get install git) before the Python environment initializes, ensuring a successful build without bloating the final image.

2. The PDF Parsing Bottleneck

• The Challenge: Financial 10-K reports often contain scanned pages or complex layouts that break standard OCR tools.
• The Solution: Implemented a robust pypdf pre-processing layer with a "Fail-Fast" mechanism. The system validates text density before passing data to the LLM, preventing wasted API tokens on unreadable files.

3. API Rate Limiting (The "Thundering Herd")

• The Challenge: LangExtract is designed for enterprise usage (Vertex AI) and attempts parallel chunk processing. On the Gemini Free Tier, this triggered 429 Resource Exhausted errors immediately.
• The Solution: Implemented an "Extraction Window" logic in the UI. Users select specific page ranges (e.g., "Risk Factors, Pages 15-20") rather than processing the entire 100-page document at once, keeping the request volume within the 15 RPM limit.

---

System Architecture

graph TD
    User[Financial Analyst] -->|Upload 10-K PDF| UI[Streamlit Interface]
    User -->|Define Schema| UI
    
    subgraph Extraction Engine
    UI -->|Raw Text| Preprocessor[PyPDF Chuncker]
    Preprocessor -->|Context Blocks| LangExtract[LangExtract Library]
    LangExtract -->|Inference Req| Gemini[Gemini 2.5 Flash]
    Gemini -->|Structured Data| LangExtract
    end
    
    LangExtract -->|HTML Generation| Visualizer[Interactive Highlighter]
    Visualizer -->|Iframe Render| UI

Loading

Core Components

1. Frontend (Streamlit): Handles file ingestion and renders the output within a secure container.
2. Intelligence (Gemini 2.5 Flash): Selected for its 1M token context window, allowing it to hold large document sections in working memory without RAG retrieval loss.
3. Grounding (LangExtract): Maps the LLM's JSON output back to the original PDF text spans using fuzzy matching algorithms.

---

⚠️ Limitations & Constraints

This is a production-grade Proof of Concept (PoC) with specific constraints:

• Text-Only Extraction: The current pipeline extracts text. It does not parse charts, graphs, or tables (requires multimodal vision upgrade).
• API Quotas: The Live Demo operates on the Google Gemini Free Tier. Heavy usage may trigger temporary cooldowns (429 Errors).
• Session State: For privacy, all files are processed in ephemeral memory and discarded immediately after the session ends. No data is persisted.

---

🚀 Installation & Local Development

Prerequisites

• Python 3.10+
• Google Gemini API Key

Setup

# 1. Clone the repository
git clone https://github.com/eatosin/DocuTrace-AI-Auditor.git
cd DocuTrace-AI-Auditor

# 2. Install dependencies (requires Git installed)
pip install -r requirements.txt

# 3. Configure Environment
export GEMINI_API_KEY="your_key_here"

# 4. Run the App
streamlit run app.py

---

Author

Owadokun Tosin Tobi Senior AI Engineer | Specialist in MLOps & LLM Evaluation

• Portfolio: ReasonBench, Sentinel
• Connect: LinkedIn

---

Built with Python, Google Cloud AI, and Engineering Rigor.