Runjie Luo
Back to Projects

AuditFlow

AI-powered Audit Intelligence & Reliability Research

Active Development
Python
FastAPI
LangGraph
PGVector
Docker
DeepSeek
PostgreSQL
Redis
Next.js
MinIO

Overview

AuditFlow is an AI audit intelligence platform exploring reliable agent workflows. It combines document processing, RAG-based retrieval, evidence grounding, and multi-agent orchestration — and the Code Archaeology series documents how its design, implementation, and runtime diverge.

Development Timeline

v0.12025-Q1

Document Parser

v0.22025-Q2

Embedding Pipeline

v0.32025-Q3

Knowledge Agent

v0.42025-Q4

Workflow Engine

v0.52026-Q1

Frontend Dashboard

v1.02026-Q2

Production Release

Features

Multi-format Document Parsing

Support PDF, Word, Excel, and scanned documents with OCR processing.

Intelligent Chunking

Semantic-aware document splitting with configurable strategies.

Vector Search & RAG

Hybrid search combining dense and sparse retrieval for accurate information lookup.

Multi-Agent Workflows

LangGraph-powered agent orchestration for complex audit procedures.

Real-time Dashboard

Interactive frontend for monitoring audit progress and results.

Knowledge Graph

Entity extraction and relationship mapping across documents.

Engineering Challenges

Problem: PDF parsing accuracy for complex layouts

Solution

Implemented a hybrid parser combining PyMuPDF for text extraction and LayoutLM for layout understanding.

Lesson Learned

Hybrid approaches significantly outperform single-parser solutions on diverse document types.

Problem: Large document processing latency

Solution

Designed a streaming pipeline with Redis-based caching and parallel chunk processing.

Lesson Learned

Streaming architecture with proper backpressure handling is critical for production document pipelines.

Links