Financial Report Intelligence System (FRIS)

An AI-powered financial intelligence platform that transforms financial reports into structured metrics, financial insights, risk assessments, and business intelligence using a multi-agent architecture.

Overview

The Financial Report Intelligence System (FRIS) is an end-to-end financial analysis platform designed to automate the extraction, validation, interpretation, and presentation of information contained in corporate financial reports.

Rather than functioning as a simple chatbot, FRIS operates as an intelligent financial analysis framework. It combines financial domain knowledge, data engineering, artificial intelligence, and software engineering to convert unstructured financial documents into actionable business intelligence.

The platform is being developed as a flagship portfolio project demonstrating modern AI engineering techniques applied to financial analysis.

Project Vision

Financial analysts spend hours manually reviewing annual reports, calculating financial ratios, identifying risks, preparing management commentary, and building dashboards.

FRIS aims to automate much of this workflow while maintaining transparency, traceability, and professional financial analysis standards.

The long-term objective is to build an extensible financial intelligence platform capable of supporting:

Financial statement analysis
Executive reporting
FP&A workflows
Investment research
Board reporting
Risk assessment
Business intelligence

Core Objectives

The system is designed to:

Extract structured financial information from reports
Automatically calculate key financial metrics and ratios
Identify trends and year-on-year movements
Detect anomalies and financial risks
Generate evidence-based management commentary
Produce dashboard specifications for Power BI
Export structured data for downstream analytics
Benchmark different open-source language models on financial reasoning tasks

System Architecture

Multi-Agent Framework

The platform is organised into specialised agents, each responsible for a focused financial task.

Order	Agent	Package	Responsibility
1	Document Processing Agent	`agent_01_document_processing`	Extract text, tables, and financial statements from PDF, Excel, and PowerPoint files
2	Metrics Extraction Agent	`agent_02_metrics_extraction`	Identify and structure financial metrics from reports
3	Financial Calculation Engine	`agent_03_financial_calculation_engine`	Calculate financial ratios, KPIs, and trends using Python
4	Data Quality Agent	`agent_04_data_quality`	Validate extracted information and detect inconsistencies
5	Financial Analysis Agent	`agent_05_financial_analysis`	Analyse trends, variances, and business performance
6	Insight Generation Agent	`agent_06_insight_generation`	Explain the business meaning behind financial movements
7	Risk Assessment Agent	`agent_07_risk_assessment`	Identify financial, operational, and reporting risks
8	Visual Blueprint Agent	`agent_08_visual_blueprint`	Generate Power BI dashboard specifications and semantic models
9	Narrative Synthesis Agent	`agent_09_narrative_synthesis`	Produce executive summaries and management reports
10	Export Agent	`agent_10_export`	Export results to JSON, CSV, Excel, Power BI-ready tables, PDF, and other destinations

Technology Stack

Backend

Python

Artificial Intelligence

Ollama (local LLM runtime)
Open-source language models
Hugging Face models
MLX (Apple silicon)

Data Engineering

Pandas
NumPy
PyMuPDF
OpenPyXL

Databases

SQLite
PostgreSQL
ChromaDB

Frontend

Streamlit

Business Intelligence

Power BI

Version Control

Git and GitHub

Model-Agnostic Design

One of the primary design goals is model independence. The Financial Intelligence Framework is not tied to any single language model, allowing supported models to be swapped without changing the core business logic.

Supported providers may include:

Ollama
Hugging Face
MLX
OpenAI
Anthropic Claude
Google Gemini
Future open-source models

This architecture enables multiple language models to be benchmarked using identical financial analysis workflows.

Planned Features

Financial Metrics Extraction

Income statement
Balance sheet
Cash flow statement
Notes to the accounts

Financial Analysis

Profitability analysis
Liquidity analysis
Solvency analysis
Efficiency analysis
Cash flow analysis
Trend analysis
Variance analysis

Automated Calculations

Gross margin
Operating margin
Net margin
EBITDA margin
Current ratio
Quick ratio
Debt-to-equity ratio
Return on assets (ROA)
Return on equity (ROE)
Interest coverage
Asset turnover
Free cash flow
Earnings per share (EPS)
Working capital

AI-Generated Insights

Executive summaries
Management commentary
Financial highlights
Risk commentary
Business drivers
Investment considerations

Export Formats

JSON
CSV
Excel
PDF
Markdown
Power BI-ready tables

Quick Start

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r development/requirements.txt
streamlit run app.py

Run the automated checks with pytest. Version 1 currently supports text-based PDFs; corrupt or image-only PDFs automatically use local Tesseract OCR. Install the OCR runtime with brew install tesseract on macOS or through your operating system's package manager.

Version 1 Capabilities

Reporting-period extraction for multi-column statements
Currency and unit-scale recognition (units, thousands, millions, and billions)
Complete structured income statement, balance sheet, and cash flow statement rows, including labels not previously known to the system
Page-level evidence for every extracted row
Two-pass OCR: low-resolution page classification followed by detailed statement OCR
Accounting equation, gross-profit, and current-ratio validation
Multi-period margins, liquidity, debt, efficiency, return, cash flow, and EPS metrics
JSON export of statements, calculations, validations, evidence, and warnings
Optional local GLM-OCR table reconstruction through Ollama
Automatic, model-assisted, and rules-only extraction modes
Per-row extraction method and confidence for model provenance
Stable 21-metric Key Financial Facts pack for cross-company analysis
Explicit missing-fact reasons instead of guessed or zero-filled values
Optional full-statement view for inspection and reconciliation
Adjacent-period amount, percentage, margin, liquidity, leverage, and cash-flow movements
Deterministic favorable, adverse, stable, turnaround, deterioration, and contextual assessments
Evidence-linked profitability, liquidity, solvency, cash-flow, performance, and reporting risk flags
Transparent severity thresholds, observed trigger values, implications, and suggested review actions
Ranked, evidence-linked management insights explaining movements and their financial meaning
Investigation prompts that avoid presenting unverified business drivers as facts
Optional local phi3:mini consolidation into connected management themes, with llama3.1:8b selectable on higher-memory or slower workflows
Post-model verification of figures, metrics, risk codes, evidence pages, and causal language

Extraction Development Checkpoint

The current extraction pipeline uses PyMuPDF, Tesseract OCR, deterministic Python table parsing, canonical financial mappings, and validation rules. Complete-row extraction has been tested against Apple, Amazon, and Colgate reports. The next extraction milestone is to add model-assisted visual table reconstruction for unfamiliar or ambiguous layouts.

Recommended models

Model	Proposed responsibility	Deployment
GLM-OCR	Primary visual table recognition, layout reconstruction, OCR, and structured row extraction	Local through Ollama
Qwen3-VL 8B	Optional financial label interpretation, period alignment, canonical mapping, and visual verification	Hosted GPU or a higher-memory development machine

If only one model is introduced initially, start with GLM-OCR because the immediate problem is accurate document and table reconstruction. Add Qwen3-VL 8B as the semantic verification layer after the GLM-OCR integration is stable.

Install the primary local extraction model with:

ollama pull glm-ocr:latest

Qwen3-VL 8B is not recommended for the current 8 GB development machine. It remains an optional future verification model that can be hosted remotely or used on higher-memory hardware after the GLM-OCR integration is stable.

Proposed hybrid extraction flow

PDF page
  → embedded text or Tesseract OCR
  → deterministic table parser
  → completeness and confidence assessment
  → GLM-OCR visual table reconstruction when required
  → Qwen3-VL financial interpretation and alignment when required
  → deterministic Python calculations
  → Data Quality Agent validation
  → human review when validation fails

Models must not silently replace validated figures or perform authoritative calculations. Every model-assisted row must retain its exact source label, statement section, period, numeric value, currency, unit, PDF page, evidence, extraction method, and confidence score.

Model-assisted extraction checklist

Local AI Model Milestone - 2 August 2026

The first model required for model-assisted extraction is now installed and verified on the development machine. This moves the project from model selection into implementation and benchmarking.

Verified environment

Component	Confirmed configuration
Development computer	Apple M2 with 8 GB unified memory
Model storage	External `Bintu2TB07` drive through the existing Ollama model directory
Ollama model	`glm-ocr:latest` (`6effedd0dc8a`)
GLM-OCR size	2.2 GB, F16
GLM-OCR architecture	1.1B parameters, 131,072-token context
Capabilities	Vision, completion, and tools
Inference	Local Apple Metal, one financial-statement page at a time

The installed supporting text models are phi3:mini, llama3.1:8b, and starcoder2:3b. They are not document-vision substitutes for GLM-OCR. The intended local allocation is:

GLM-OCR: visual document, table, and statement extraction.
Phi-3 Mini: lightweight explanations and narrative synthesis from validated JSON.
Llama 3.1 8B: optional deeper text analysis, run separately because of memory limits.
StarCoder2 3B: optional Python, SQL, DAX, and dashboard-code assistance.

The application will keep deterministic Python calculations and accounting validation as the authoritative layer. Language models may extract or explain evidence, but must not silently replace validated financial values.

Model integration completed

Agent 01 now includes an Ollama/GLM-OCR provider, health checks, sequential page rendering, HTML/Markdown table normalization, provenance, and safe deterministic fallback. Streamlit offers automatic, model-assisted, and rules-only modes and displays the model and extraction method used. Model-assisted figures are routed through the existing structured-statement schema, calculations, and accounting validation before display or export.

The next implementation milestone is to benchmark live GLM-OCR output against Apple, Amazon, and Colgate, tune the page prompts and table normalization where required, and record deterministic-versus-model row accuracy and processing time.

Insight Intelligence Milestone - 3 August 2026

Agents 05, 06, and 07 now form a traceable financial interpretation workflow after statement extraction and calculation:

Validated facts and ratios
  → adjacent-period trend and variance analysis
  → deterministic financial risk assessment
  → evidence-grounded insight generation
  → optional local-model management-theme synthesis
  → post-model numerical and provenance verification

Agent 05 classifies multi-period movements as favorable, adverse, stable, turnaround, deterioration, or contextual. Agent 07 applies transparent profitability, liquidity, solvency, cash-flow, performance, and reporting-quality risk thresholds. Agent 06 ranks the resulting insights and groups them into profitability, cash-flow, liquidity, leverage, and shareholder- return themes before any language model is called.

The optional narrative layer uses phi3:mini by default on the 8 GB M2. A live Amazon 2025 test produced five verified management themes covering cash generation and investment, capital structure and leverage, liquidity and working capital, growth and profitability, and shareholder returns and efficiency. The model connected operating cash-flow growth with the free-cash-flow decline and balanced debt growth against the change in debt-to-equity.

Two live llama3.1:8b attempts exceeded the practical local timeout, so it remains selectable for higher-memory or deliberately slower workflows rather than being the default. Model output cannot change calculations, priorities, metrics, risks, or evidence. Python rejects invented figures and unsupported causal claims, restores any omitted validated metric sentence, strips unverified currency symbols, and retains deterministic insights whenever augmentation is unavailable or invalid. The complete automated suite contains 37 passing tests.

Development Update - 2 August 2026

Today’s work strengthened the financial-statement extraction foundation and made the project easier to continue in a future development session.

Completed today

Replaced the limited KPI allowlist as the primary extractor with complete-row statement extraction that preserves exact PDF labels, sections, periods, values, ordering, and evidence.
Corrected cash-flow and balance-sheet extraction across different annual-report layouts.
Added Amazon and Colgate annual reports alongside Apple as regression samples.
Visually compared extracted tables with rendered source pages.
Expanded deterministic validation to include balance-sheet totals, cash rollforwards, and operating + investing + financing + foreign-exchange reconciliations.
Improved Streamlit statement inspection with exact labels, section names, row counts, formatted amounts, larger tables, and consistent ratio formatting.
Numbered the agent packages from agent_01_... through agent_10_... so their intended workflow order is visible in the repository.
Documented the next model-assisted extraction milestone using GLM-OCR for visual table reconstruction and Qwen3-VL 8B for financial interpretation and verification.

Validation snapshot

Report	Income rows	Balance-sheet rows	Cash-flow rows	Reconciliations
Apple	19	27	33	12/12 passed
Amazon	23	26	29	10/10 passed
Colgate	17	31	33	15/15 passed

The automated suite now contains 37 passing tests, including provider, table-normalization, page-rendering, provenance, and safe-fallback coverage. Live Colgate model-assisted extraction processed the three primary statements in 419 seconds on the 8 GB M2 and retained 17 income, 31 balance-sheet, and 33 cash-flow rows with all 15 validations passing.

Milestone: Phase 1 Complete

Phase 1 was completed and validated against the included Apple, Amazon, and Colgate annual reports. The Apple pipeline automatically detected corrupt embedded PDF text, classified the document at low resolution, and applied detailed OCR only to the primary statements.

Verified extraction

Statement	Source page	Reporting periods	Currency and unit
Consolidated statements of operations	32	2022, 2021, 2020	USD millions
Consolidated balance sheets	34	2022, 2021	USD millions
Consolidated statements of cash flows	36	2022, 2021, 2020	USD millions

Verified 2022 calculations

KPI	Result
Gross margin	43.31%
Operating margin	30.29%
Net margin	25.31%
Current ratio	0.88x
Working capital	$(18,577) million
Debt-to-equity	2.37x
Debt ratio	0.34x
Asset turnover	1.12x
Return on average assets	28.36%
Return on average equity	175.46%
Free cash flow	$111,443 million
Basic / diluted EPS	$6.15 / $6.11

Cross-report extraction coverage

Sample	Income rows	Balance-sheet rows	Cash-flow rows	Checks passed
Apple 2022 Form 10-K	19	27	33	12/12
Amazon 2025 Annual Report	23	26	29	10/10
Colgate 2025 Annual Report	17	31	33	15/15

An external EE Limited FY2015 IFRS report was also validated against the Key Financial Facts workflow. The parser identified Group income statement, Statements of financial position, and the two-page Statements of cash flows; selected Group rather than Company columns; and returned 17/21 canonical facts in approximately 0.35 seconds without model assistance. Missing facts remained explicit rather than being inferred from narrative notes.

The complete automated suite contains 37 passing tests. Full two-pass Apple OCR and analysis completes in approximately 48 seconds on the development machine; text-native Amazon and Colgate reports complete in approximately one second each.

Agent implementation status

Agent	Status
Document Processing Agent	Implemented
Metrics Extraction Agent	Implemented
Financial Calculation Engine	Implemented
Data Quality Agent	Implemented
Financial Analysis Agent	Implemented
Insight Generation Agent	Implemented
Risk Assessment Agent	Implemented
Visual Blueprint Agent	Planned folder ready
Narrative Synthesis Agent	Baseline implemented
Export Agent	JSON available; additional formats planned

Roadmap

Version 1

PDF upload ✅
Multi-period structured statement extraction ✅
Financial ratio and KPI calculation ✅
Financial equation validation ✅
Targeted OCR fallback ✅
Executive summary ✅
Multi-period trend and variance analysis ✅
Evidence-linked financial risk assessment ✅
Evidence-grounded management insight generation ✅
JSON export ✅

Version 2

Multi-agent workflow
Risk detection
Dashboard blueprint generation
Power BI export
Financial data validation

Version 3

Multi-year trend analysis
Peer comparison
Company benchmarking
Investment scoring
Stock valuation models
Model benchmarking framework

Version 4

Live financial data integration
Portfolio analysis
Intelligent research reports
Financial knowledge graph
Personal AI financial research assistant

Repository Structure

financial-report-intelligence-system/
├── backend/fris/                  # Application package
│   ├── agents/                    # One folder per agent responsibility
│   │   ├── agent_01_document_processing/
│   │   ├── agent_02_metrics_extraction/
│   │   ├── agent_03_financial_calculation_engine/
│   │   ├── agent_04_data_quality/
│   │   ├── agent_05_financial_analysis/
│   │   ├── agent_06_insight_generation/
│   │   ├── agent_07_risk_assessment/
│   │   ├── agent_08_visual_blueprint/
│   │   ├── agent_09_narrative_synthesis/
│   │   └── agent_10_export/
│   ├── models.py
│   └── pipeline.py
├── development/
│   ├── tests/                     # Automated tests
│   ├── tools/                     # Developer and maintenance tools
│   └── requirements.txt           # Development dependencies
├── docs/
│   ├── assets/
│   └── architecture.md
├── sample_reports/                # Local reports used for evaluation
├── app.py                         # Streamlit entry point
├── pyproject.toml                 # Python package configuration
├── requirements.txt               # Runtime installation
└── README.md

Development-only tests, dependencies, and tools are consolidated under development/. Every responsibility in the Multi-Agent Framework has its own folder under backend/fris/agents/, including documented placeholders for planned agents. See each folder's README.md and docs/architecture.md for its status and extension boundary.

Guiding Principles

Financial accuracy before AI creativity
Python performs deterministic calculations
AI provides interpretation and reasoning
Every insight should be traceable to source evidence
Modular architecture with interchangeable models
Transparent and explainable financial intelligence

Disclaimer

This project is intended for financial analysis, research, education, and decision support. It does not provide financial advice or investment recommendations. Users remain responsible for validating outputs and making their own financial decisions.

Author

Teslim Adeyanju, ACA
Financial Data Analyst | Chartered Accountant | AI & Financial Intelligence Engineer

Building practical AI systems that combine finance, data engineering, and machine learning to solve real-world business problems.

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
backend/fris		backend/fris
development		development
docs		docs
sample_reports		sample_reports
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
app.py		app.py
pyproject.toml		pyproject.toml
requirements.txt		requirements.txt

Latest commit

History

Folders and files

Repository files navigation