TaxForward Gemini

Inspiration

Most tax tools answer a household's question after the useful decision window has closed. TaxForward starts from a different question: what can I still decide before December 31, and is a tax benefit really worth the cash cost, risk and loss of liquidity? The idea comes from Michel Chatel's 26 years of accounting and tax experience. The aim is to make reliable tax decision support understandable and affordable for standard French households.

What it does

TaxForward guides a standard metropolitan-French household through composition, income, professional expenses, financial income, standard micro-property and micro-BIC cases, deductions, credits and withholding. A versioned deterministic engine projects 2027 tax from estimated 2026 income, compares PFU and progressive-scale combinations, and creates decision cards for PER retirement savings, donations, home services, professional expenses and eligible investment scenarios.

Every card shows the possible tax effect beside cash required, net effort, deadline, liquidity, risk and eligibility checks. Gemini then explains the already-computed projection and the two most relevant decisions in plain French. Gemini never calculates the tax and cannot modify the engine's results.

How we built it

The application uses React 19, TypeScript, a Next.js-compatible App Router on Vinext, a versioned France 2026 rule pack and Vitest. The interface calls a protected server endpoint only after the deterministic engine has produced the projection and decision cards. The endpoint uses the Google GenAI SDK and the Gemini Developer API.

Before the model call, the server creates a minimized payload containing computed results, decision cards and limited risk and liquidity context. Raw salary fields and the full household profile are excluded. The system instruction forbids recalculation, derived values and product promotion. After generation, a numeric guard verifies every euro amount and percentage against the transmitted payload. Unsupported claims cause an automatic deterministic fallback.

Why Gemini

The difficult part is not only calculating tax. It is explaining the trade-off between tax saving, cash commitment, liquidity and risk. Gemini is used where a language model adds value: clear, contextual explanation. The deterministic engine remains the single source of truth.

This boundary is visible in production. The calculation is fixed before Gemini is called. The interface identifies whether an explanation came from Gemini or from the deterministic fallback. The tax engine never consumes Gemini output.

Challenges we ran into

The central challenge was keeping probabilistic language completely outside an auditable tax computation. Prompt instructions alone were not sufficient, so we added payload minimization and an output guard that rejects unsupported numeric claims. During production testing, Gemini formatted 7,657 euros with a different thousands separator from the engine. We updated the guard to recognize equivalent formatting while still rejecting genuinely invented amounts.

A second challenge was presenting tax optimization honestly. A lower tax bill can still be a poor decision when capital is locked, fees are high or loss is possible. The decision cards therefore show net effort, timing, liquidity and risk alongside the potential tax effect.

Accomplishments that we are proud of

• A functioning consumer flow, not a calculation-only proof of concept.
• A clear architecture: deterministic calculation first, Gemini explanation last.
• A one-click synthetic judge demo with no identity or tax number required.
• PFU and progressive-scale comparison, including an independent crypto choice.
• Person-specific professional expenses and PER ceilings.
• A tested numeric safety boundary and deterministic fallback.
• 35 passing automated tests and a successful production build.
• A live Gemini explanation verified in the public competition deployment.

What we learned

Generative AI is most trustworthy in a regulated calculation product when its role is narrow, observable and reversible. The strongest design is not "AI calculates everything". It is "auditable software calculates, and AI makes the verified result understandable".

What's next

Next steps are to expand the reviewed French rule pack, add consent-based tax-notice import, automate legal-change diffs, improve model-output provenance, obtain independent tax and legal review, and add separate country-year packs without mixing jurisdictions.

Limitations

TaxForward is a competition prototype, not an official tax calculation or individualized financial advice. It handles stable, standard metropolitan-French situations. Event years, non-residents, overseas rules, foreign income, real-property regimes and other complex cases are excluded or deferred. The 2027 Finance Act is not final. No financial product is recommended or sold.

Built With

• ai
• gemini
• genai
• github
• google
• openai
• typescript