PinSight: AI-Driven Audit Execution Platform

Deloitte China2025.05
AI Audit
Product Architecture
Document AI
Workflow Automation
Digital Transformation

Project Overview

PinSight is an AI-driven audit execution platform designed for SME audit teams, addressing the inefficiencies of traditional labor-intensive auditing. As enterprise data complexity grows, manual vouching and rule-based sampling become unreliable and error-prone. Meanwhile, most SME audit teams lack access to expensive enterprise systems, creating a critical “digital gap”. PinSight bridges this gap by automating the full audit workflow — from risk-based sampling and document understanding to logic validation and working paper generation — through a modular, explainable AI system. The system emphasizes: - Precision over random sampling - Structured understanding of unstructured audit evidence - Human-in-the-loop control for reliability and compliance This project was delivered as a working product prototype with a functional UI and end-to-end workflow.

Key Innovations

  • Risk-oriented smart sampling using hybrid ML models (XGBoost + Isolation Forest) to detect anomalies beyond traditional methods.
  • OCR-free document understanding via Donut (Document Transformer) for direct image-to-JSON extraction.
  • Triple-check validation engine covering consistency, logic, and cross-document correlation.
  • Automated audit narrative generation using LLMs with traceable reasoning.
  • Modular microservices architecture enabling plug-and-play deployment.

System Workflow

Import: Upload ledgers or datasetsSample: AI identifies high-risk entriesRecognize: Extract structured data from vouchersValidate: Run business logic checksGenerate: Produce audit workpapers

What I Did

  • Led product development as the sole technologist, translating audit workflows into an AI-driven system.
  • Designed the full architecture: ingestion, sampling, document understanding, validation, and reporting.
  • Built a working web prototype with interactive UI for anomaly inspection and report generation.
  • Developed prompt workflows for AI-assisted audit reasoning and narrative generation.
  • Prioritised explainability and auditor control in system design.
  • Delivered MVP under time constraints focusing on usability and reliability.

Reflection

This project reinforced a key insight: in audit systems, reliability and trust matter more than raw model performance. Instead of building a black-box AI, I designed a system where auditors can verify, control, and override every step. The modular architecture ensures flexibility and real-world deployability, especially for SME audit environments. Being the sole technical contributor also strengthened my ability to translate domain knowledge into executable product systems.