Veritas AI

Inspiration: The Crisis of Verification

We live in an era of information abundance but verification scarcity. Financial systems—the operating system of our global economy—have become so complex that human auditors struggle to effectively validate them. A single error or omission can cascade into catastrophic risks, yet State-of-the-Art (SoA) practice still relies on manual cross-referencing and "gut checks."

We built Veritas AI not merely to automate financial report reviews, but to explore a fundamental question in AI safety: Can we build a Multi-Agent System (MAS) that cares about truth more than completion?

This project is the first step in our journey toward building MAS systems for the benefit and safety of humanity. We set out to build a cognitive architecture that scrutinizes financial reports—checking the math, debating the logic, and verifying claims against the real world. Surprisingly, in just three weeks, we built a system using Gemini 3 Pro that performs significantly better than current industry solutions in both accuracy and speed.

What it does

Veritas AI is an autonomous "Co-Auditor" that significantly increases the efficiency of pre-sign-off financial statement reviews. It replaces days of manual checking with minutes of agentic processing across four key pillars:

Numeric Validation: Performs "Tick and Tie" procedures, ensuring numbers reconcile within tables and flow correctly across different financial statements (e.g., Income Statement to Cash Flow).
Logic & Consistency: Checks for semantic contradictions (e.g., text says "revenue up" while data shows "revenue down") and business logic violations.
Compliance (IFRS/IAS): Dynamically identifies applicable accounting standards and verifies disclosure requirements.
External Signal Verification: Uses Gemini Deep Research to cross-reference report claims with external authoritative sources (Internet -> Report) and checks if external material events are properly disclosed (Report -> Internet).

How we built it: The Gemini 3 Advantage

We utilized the full Google ecosystem to create a Neuro-Symbolic Architecture (combining probabilistic reasoning with deterministic execution):

Gemini 3 Pro (The Brain): Used for high-level reasoning, cross-table logic detection, and the "Audit Orchestrator." Its expanded context window and reasoning capabilities were essential for the "3x3 Multi-Pass" logic.
Gemini Flash: Used for high-speed, lower-complexity tasks like initial document classification and table naming.
Google Agent Development Kit (ADK): The backbone of our MAS efficiency, allowing us to manage state and handoffs between the Orchestrator and sub-agents.
Google Antigravity: Leveraged for rapid coding efficiency to build the fullstack execution layer.
Gemini Deep Research: The engine behind our "External Signal" agent, allowing the system to autonomously browse the web and synthesize multi-source findings.

Key Architectural Patterns

Deterministic + LLM Hybrid (The Anchor Formula): We operate on the rule: LLMs for inference, Python for calculation. The LLM analyzes the first row of a table to deduce the formula, then hands it to a Python engine to scale the calculation across the dataset. This ensures 100% arithmetic accuracy with zero hallucinations.
3x3 Exploration => Verification: For deep logical problems, we implemented a pipeline where agents explore multiple lines of reasoning (Exploration) and then rigorously debate their own findings (Verification) before finalizing a result.
Thought Signatures: We implemented sequential pipelines where agents "sign" their thought process, allowing downstream agents to audit the reasoning of upstream agents.

Challenges we ran into

1. Scaling Table Footings (The Math Problem) Our initial attempts to have the LLM check every cell failed due to context limits and latency.

The Solution: We realized we only needed the LLM to understand the logic of the table. We developed an approach where Gemini derives the formula from the first cell, and we use deterministic programming to apply it to the rest. Combined with optimal table segmentation, we cracked the scaling problem.

2. Deep Logical Reasoning (The Expertise Problem) Building a system that could catch subtle business logic errors (which usually requires many years of audit experience) seemed intractable. We found that models would find 1-2 deep issues out of 20, but not consistently.

The Solution: We combined three insights:
1. Models can find deep issues, just not all at once.
2. Models are probabilistic (different turns yield different insights).
3. Single-pass results are capped. This led to the revelation of the 3x3 Exploration => Verification/Exploitation system. By running 3 parallel chains through 3 refinement passes, we achieved results that shocked us—uncovering deep logical inconsistencies that even human experts might miss.

Accomplishments that we're proud of

SoA Performance: In 3 weeks, we built a system that outperforms current state-of-the-art manual and automated solutions.
Architectural Breakthroughs: The "Anchor Formula" pattern and "3x3 Refinement" loops are reusable architectures for any domain requiring high-precision validation.
Mission Alignment: We successfully demonstrated that an AI agent can be engineered to prioritize truth over task completion, a critical step for AI safety.

What we learned

The Limits of Gemini 3: We learned exactly where the model's reasoning hits a wall and how to use agentic collaboration to climb over that wall.
Cooperation is Key: The quality of the output is defined not just by the model, but by how effectively the agents pass context and "debate" one another.
Demo Early: We learned the hard way that user feedback is vital; we shifted from pure feature development to a deployable demo to validate our assumptions.

What's next for Veritas AI

Perfection: We will refine the "Accounting Lens" agents to handle even more niche IFRS standards.
Scaling: Moving from a prototype to a scalable MaSS (multi agent system as a service) capable of handling the global load of financial reporting.
The Long Mission: This is just the beginning. We aim to apply this "Truth-Seeking" MAS architecture to other critical domains like economic and statistical reports, medical records, and scientific research.