My Contributions

Three major feature systems added to the open-source Ghostfolio wealth management platform.

AI Financial Education Agent

Interactive chat panel that answers financial questions using the user's actual portfolio data. Designed as an educational tool — explains concepts, never gives investment advice.

7 Specialized Tools

Glossary lookup — financial term definitions
Metric explainer — Ghostfolio metrics in context
Compound interest calculator
Concept comparison — side-by-side education
Currency conversion explainer
Portfolio summary — snapshot overview
State tax lookup — jurisdiction-specific rules

3 Safety Interceptors

Advice Boundary — regex-based filtering prevents investment advice
Hallucination Detection — flags claims not grounded in tool results
Output Validation — ensures response structure and completeness

US Tax Module

Production-grade capital gains reporting that follows IRS methodology. Pure, deterministic computation — the LLM routes and narrates, but every number comes from tested functions.

FIFO Capital Gains — Deterministic lot matching with fee proration and FX conversion. 42 unit tests on the pure engine.

IRS Treasury FX Rates — Quarterly rates from the Treasury API — matches IRS methodology, avoids market-rate discrepancies.

Wash Sale Detection — Best-effort 30-day window flagging. Detects suspicious patterns; users apply judgment for edge cases.

State Tax Estimation — State-specific income tax rules with selector UI and estimate cards.

CSV Export — Tax lot data exported in a format ready for tax filing software.

Translation Layer

20+ language support using a tiered strategy:

Tier 1: LLM-Native

12 languages the LLM handles natively — no translation step needed. English, Spanish, French, German, etc.

Tier 2: NLLB Translation

200+ languages via Meta's NLLB neural translation model. Self-hosted on a dedicated GPU with RTL layout support for Arabic, Hebrew, Farsi.

Eval Suite

200+ test scenarios across 9 categories with 15 automated scorers (4 deterministic, 11 LLM-judged). CI pipeline runs evals on every change to AI education code.

Scorer	Score
hasDisclaimer	100%
safetyCompliance	100%
relevance	97%
noInvestmentAdvice	95%