FinSherlock

An AI-powered financial investigation platform designed to help everyday investors understand financial statements, identify potential accounting red flags and make more informed investment decisions.

Comment

1. Project Title

FinSherlock: AI Financial Investigator

Turning Financial Filings into Evidence-Backed Insights

Theme: Artificial Intelligence & Machine Learning

2. Project Description

Problem Statement

The rapid growth of retail investing has made financial markets accessible to millions of people, but understanding financial information remains difficult.

Companies publish extensive financial disclosures such as 10-Ks, 10-Qs, 8-Ks, annual reports, and other filings, often running hundreds of pages. These documents contain critical information about a company's financial health, accounting practices, liabilities, and potential risks but are primarily written for professionals with accounting and financial-analysis expertise.

As a result, many retail investors rely heavily on headlines, stock prices, social-media sentiment, or simplified metrics such as EPS. Important warning signs can remain hidden in footnotes, MD&A sections, accounting policies, and relationships between different financial metrics.

Existing solutions do not completely solve this problem:

  • Stock screeners provide numbers but limited context.
  • Financial media and online influencers can be biased or difficult to independently verify.
  • Generic AI chatbots can generate fluent answers but may hallucinate or provide answers without evidence.
  • Professional forensic-accounting software is often expensive, complex, and designed for auditors.

The problem is therefore not a lack of financial data it is the lack of an accessible, explainable, evidence-backed AI system that can investigate financial disclosures for everyday users.

Solution Overview

FinSherlock is an AI-powered financial investigation platform that transforms complex financial filings into an interactive investigation.

Users upload a company's financial report, and our AI agent Sherry investigates the document using a domain-specific financial-risk taxonomy and a Retrieval-Augmented Generation (RAG) pipeline.

Instead of simply summarizing the document, Sherry:

  1. Reads and processes the financial filing.
  2. Identifies potentially suspicious financial patterns.
  3. Retrieves the relevant evidence from the original document.
  4. Assigns severity to identified risks.
  5. Explains the finding in simple, non-technical language.
  6. Provides the exact source passage and page reference.
  7. Allows the user to interrogate the evidence conversationally.

This makes FinSherlock an AI financial investigator rather than a generic financial chatbot.

The core principle is:

Don't just give the user an answer. Give them the evidence behind the answer.

This evidence-first architecture is designed to reduce hallucination risk and make AI-generated financial analysis more transparent and auditable.

AI/ML Innovation

FinSherlock's primary innovation lies in combining Generative AI, semantic retrieval, document intelligence, and domain-specific financial reasoning.

Rather than giving an LLM unrestricted access to a document and asking it to "find fraud," we created a structured investigation pipeline.

Financial Risk Taxonomy

Sherry is trained to investigate six core categories:

  • Revenue Recognition Inconsistencies
  • Cash Flow vs. Earnings Mismatches
  • Accounting Pattern Shifts
  • Unusual Related-Party Transactions
  • Aggressive Financial Forecasting
  • Off-Balance-Sheet Arrangements

Each finding receives a severity level and supporting evidence.

Retrieval-Augmented Generation

The RAG architecture ensures that Sherry's answers are grounded in the actual financial filing.

The pipeline is:

PDF → Text Extraction → Chunking → Embeddings → Vector Database → Semantic Retrieval → Risk Analysis → LLM Reasoning → Evidence-Linked Finding

This approach allows Sherry to retrieve the most relevant sections of a long document before generating an explanation.

Key Features

1. AI Financial Investigator — Sherry

Sherry acts as an AI financial detective that investigates company filings and communicates findings in plain English.

2. Intelligent Financial Document Analysis

Users can upload:

  • 10-K
  • 10-Q
  • 8-K
  • Annual Reports
  • Other financial/SEC filings

The system processes the document automatically rather than requiring users to manually search hundreds of pages.

3. Automated Red-Flag Detection

FinSherlock identifies potentially concerning financial patterns using its structured risk taxonomy.

4. Evidence-Backed AI

Every major finding is accompanied by:

  • Severity score
  • Supporting evidence
  • Source passage
  • Page reference

This allows users to verify the AI's reasoning against the original document.

5. Conversational Investigation

Users can ask Sherry follow-up questions such as:

  • "Why is this considered a risk?"
  • "Show me the evidence."
  • "What happened to operating cash flow?"
  • "Has management mentioned this before?"

Sherry retrieves the relevant source material and responds in context.

6. Investigation Dashboard

The interface presents:

  • Overall risk score
  • Red flags
  • Evidence cards
  • Charts
  • Investigation timeline
  • Source references
  • Interactive conversation with Sherry

7. Rapid Analysis

Instead of manually spending hours reviewing a lengthy financial filing, users receive preliminary investigative findings within minutes.

3. Target Users

Retail & Beginner Investors

People who want to understand what they are investing in but find financial statements intimidating.

DIY Investors

Users who already read financial filings but lack professional forensic-accounting training.

Financial Students & Researchers

Students and researchers can use FinSherlock as an interactive tool for exploring real-world financial statements.

Analysts & Finance Professionals

FinSherlock can serve as a rapid first-pass investigation tool, helping analysts identify areas that deserve deeper manual review.

Financially Vulnerable Investors

Users with limited financial literacy can receive complex financial information in accessible, plain-English explanations.

4. Technologies Used

Frontend

  • Next.js 14
  • React 18
  • Tailwind CSS
  • Framer Motion

Used to create the interactive investigation dashboard, evidence board, ChatPanel, risk meter, and user experience.

Backend

  • FastAPI
  • Uvicorn
  • SQLite
  • SQLAlchemy
  • ChromaDB

Used for document processing, RAG orchestration, retrieval, persistence, and investigation management.

AI/ML

  • Gemini 2.5 Flash: LLM inference
  • all-MiniLM-L6-v2: semantic embeddings
  • ChromaDB: vector database
  • Cosine Similarity: semantic retrieval
  • RAG architecture: evidence-grounded generation

Document Intelligence

  • PyMuPDF: PDF extraction
  • RecursiveCharacterTextSplitter: document chunking

The complete pipeline extracts, chunks, embeds, retrieves, analyzes, and persists evidence from financial filings.

Deployment

  • Vercel: frontend
  • Hugging Face Spaces
  • Docker: backend deployment

5. Technical Architecture

                 USER
                   │
                   ▼
          ┌─────────────────┐
          │  FinSherlock UI │
          │ Next.js + React │
          └────────┬────────┘
                   │
                   ▼
          ┌─────────────────┐
          │     FastAPI     │
          │     Backend     │
          └────────┬────────┘
                   │
                   ▼
          ┌─────────────────┐
          │   PDF Ingestion │
          │    PyMuPDF      │
          └────────┬────────┘
                   │
                   ▼
          ┌─────────────────┐
          │     Chunking    │
          └────────┬────────┘
                   │
                   ▼
          ┌─────────────────┐
          │    Embeddings   │
          │ all-MiniLM-L6-v2│
          └────────┬────────┘
                   │
                   ▼
          ┌─────────────────┐
          │    ChromaDB     │
          │  Vector Store   │
          └────────┬────────┘
                   │
                   ▼
          ┌─────────────────┐
          │ Semantic RAG    │
          │   Retrieval     │
          └────────┬────────┘
                   │
          ┌────────┴────────┐
          ▼                 ▼
 ┌─────────────────┐ ┌──────────────────┐
 │ Financial Risk  │ │ User's Question  │
 │    Taxonomy     │ │                  │
 └────────┬────────┘ └────────┬─────────┘
          └──────────┬─────────┘
                     ▼
            ┌─────────────────┐
            │  Gemini 2.5     │
            │     Flash       │
            └────────┬────────┘
                     │
                     ▼
            ┌─────────────────┐
            │ Evidence-Linked │
            │    Findings     │
            └────────┬────────┘
                     │
                     ▼
            ┌─────────────────┐
            │ Sherry + Risk   │
            │ Evidence Board  │
            └─────────────────┘

6. Why FinSherlock Fits the AI/ML Theme

FinSherlock goes beyond simply adding an LLM chatbot to a conventional application.

The project demonstrates practical AI/ML through:

  • Semantic embeddings
  • Vector search
  • Retrieval-Augmented Generation
  • Generative AI
  • Domain-specific AI reasoning
  • Document intelligence
  • AI-powered classification
  • Evidence-grounded conversational AI

The system combines these technologies into a complete workflow that solves a real-world problem: making complex financial information understandable and actionable for people who lack specialized financial expertise.

7. Creativity & Originality

The creative element of FinSherlock is the combination of financial forensics + RAG + conversational AI + detective-style UX.

Most financial applications focus on:

Price → Charts → Ratios → Buy/Sell signals

FinSherlock focuses on:

Document → Evidence → Pattern → Risk → Investigation

The "detective" metaphor is not merely cosmetic. It reflects the underlying product philosophy: financial statements contain clues, and Sherry's job is to connect those clues into an evidence-backed case.

This makes the product more engaging for beginners while maintaining technical depth for advanced users.

8. User Experience

FinSherlock is designed around a simple five-step journey:

1. Start Investigation ↓ 2. Upload Financial Report ↓ 3. Sherry Investigates ↓ 4. Review Case File & Risk Score ↓ 5. Interrogate the Evidence

The user does not need to understand accounting standards, RAG, embeddings, or vector databases.

The complexity stays behind the interface while the user sees a simple investigation workflow.

The dashboard uses an Investigation Room, Case File, Evidence Board, Risk Meter, and conversational Sherry to make an otherwise intimidating financial-analysis process intuitive.

9. Future Scope

Multi-Year Financial Forensics

Analyze multiple years of financial statements to detect trends such as:

  • Receivables growing faster than revenue
  • Margin erosion
  • Persistent cash-flow/earnings mismatches
  • Increasing liabilities
  • Repeated accounting changes

Cross-Filing Investigation

Connect 10-Ks, 10-Qs, 8-Ks, earnings releases, and historical reports to identify inconsistencies across disclosures.

Real-Time Risk Monitoring

Monitor newly published filings and notify users when significant financial-risk patterns emerge.

Advanced Forensic Models

Integrate additional quantitative financial-forensics techniques for accrual quality, revenue quality, working capital, cash conversion, and other indicators.

Broader Market Coverage

Expand beyond the initial India-focused use case to support international markets, financial reporting standards, and additional filing formats.

Professional Research Mode

Introduce an advanced mode for analysts and researchers with deeper financial metrics, customizable risk thresholds, comparative company analysis, and exportable investigation reports.

10. Expected Impact

FinSherlock aims to reduce the gap between financial information and financial understanding.

Instead of expecting every investor to become a forensic accountant, we use AI to make the first layer of investigation accessible to everyone.

Our vision is:

Every investor should be able to ask not only "What are the numbers?" but "What do these numbers actually tell me—and where is the evidence?"

FinSherlock turns hundreds of pages of financial disclosure into a structured, explainable, evidence-backed investigation, demonstrating how AI can be applied to a real-world problem with meaningful user impact.

Built With

Share this project:

Updates

Submission history