Audit Risk Forecast Tracker

Gemini 3 -powered tax audit risk forecasting platform

Comment

Inspiration

Self-filing taxes creates a specific kind of anxiety: not whether numbers add up, but whether they look suspicious from an auditor's perspective. Most bookkeeping tools focus on recording transactions. Very few help users anticipate how those records might be interpreted by tax authorities.

I wanted a lightweight system that:

  • Works directly with Google Sheets (no heavy accounting migration)
  • Does not store sensitive financial data in a separate system
  • Uses AI not for bookkeeping, but for simulating audit questions

The core inspiration was this shift:

Instead of predicting audits, simulate what might trigger one.

What it does

Audit Risk Forecast Tracker analyzes spreadsheet-based expense data and:

  1. Detects structural anomalies:

    • Extreme year-over-year changes
    • Category concentration spikes
    • Cross-category identical transactions
    • Missing receipts
    • High-amount outliers
    • Frequent transaction density

  2. Quantifies statistical signals:

    • Growth rate
    • Ratio shift
    • Category composition
    • Z-score deviations

  3. Uses Gemini to generate:

    • A "Tax Authority Perspective"
    • A Primary Focus account
    • Suggested preparation actions

Rather than telling users "you will be audited," it asks:

"If I were reviewing this, what would I question?"

The system transforms raw accounting data into a structured audit simulation.

How we built it

The project evolved in three stages.

1. Spreadsheet-linked accounting foundation

  • Google Sheets as the source of truth
  • Aggregation logic for yearly expense summaries
  • Deterministic anomaly detection logic

The application does not permanently store financial data. Data is either computed on demand or cached in summarized form.

2. Statistical signal engine

We implemented numerical analysis for:

  • Category ratio concentration
  • Year-over-year growth
  • Absolute amount thresholds
  • Cross-category matching
  • Z-score calculation: Z = (X - μ) / σ

This logic is handled deterministically in code — not by AI.

3. Gemini-powered audit simulation

Gemini is used only after numerical signals are computed. Its role:

  • Interpret structured anomaly data
  • Frame them as plausible audit inquiries
  • Suggest preparation steps

Strict separation of responsibility:

  • Code handles math
  • AI handles narrative reasoning

Frontend:

  • React + Vite
  • Firebase (optional cache)
  • Clean UI that mimics real audit questioning flow

Challenges we ran into

1. Numerical robustness

Raw percentage growth can be misleading with small baselines. Designing fair anomaly thresholds required iterative tuning.

2. AI control and hallucination prevention

Early prompts caused:

  • Overconfident conclusions
  • Statistically unjustified narratives

We refined prompts to:

  • Restrict AI to interpreting provided signals
  • Avoid generating unsupported claims

3. Data architecture decisions

We intentionally chose:

  • Spreadsheet as the single source of truth
  • No full financial replication in Firestore

This required careful separation between:

  • Live aggregation
  • Optional caching
  • Demo data handling

Accomplishments that we're proud of

  • Clear separation between deterministic analytics and AI reasoning
  • Lightweight architecture with no financial data lock-in
  • Controlled AI output aligned with statistical evidence
  • Converting a basic expense tracker into a structured audit-preparation tool

The project evolved from bookkeeping to decision-support.

What we learned

  1. AI should not perform numerical analysis blindly.
  2. Statistical signals must be explainable before they are narratively framed.
  3. Prompt engineering in compliance contexts requires strict guardrails.
  4. Simulation can be more useful than prediction.

Most importantly:

AI adds the most value when it explains structured evidence — not when it replaces it.

What's next for Audit Risk Forecast Tracker

  1. Multi-year comparative analysis dashboard
  2. Risk trend visualization over time
  3. Industry benchmark comparison
  4. Automated explanation draft generation
  5. Document management integration (receipt attachment tracking)
  6. Risk score calibration based on real-world audit datasets

Long term, this can evolve into:

A proactive compliance companion for independent professionals.

Built With

Share this project:

Updates