Fidelio

Inspiration

The average American pays $138,000 in unnecessary 401(k) fees over a lifetime—and has absolutely no idea. Plan statements are deliberately opaque, expense ratios hide behind three decimal points, and fund names sound sophisticated while quietly bleeding your returns.

We realized that this is fundamentally an information asymmetry problem. A simple 1% difference in fees doesn't just reduce your return by 1%; because of the compounding effect, it geometrically destroys your wealth over decades. Mathematically, the difference in future value ($FV$) with an annual contribution ($C$), market return ($r$), and fee drag ($f$) over $n$ years is devastating:

$$ FV = \sum_{t=1}^{n} C \times (1 + r - f)^{n-t} $$

If you run the numbers on a $50k salary over 30 years, that tiny fee variable $f$ strips away $84,500. Add in missed employer matches, and the wealth gap widens further. We wanted to build a tool that acts as an elite, uncompromised financial advocate for the everyday worker—translating complex PDFs into plain English in seconds.

What it does

Fidelio (from Latin fidelis, faithful) is an AI financial translator that exposes hidden fees and missed money.

You simply drop a 401(k) statement or paystub PDF into the app. In 30 seconds, Fidelio reads it and generates a personalized, plain-English report containing:

• Health Score & Headline: A 0–100 score and the single most important insight from your document.
• Holdings Analysis: Every fund you own, its expense ratio, and how it compares to cheaper alternatives.
• Fee Impact Projections: The exact dollar amount you will lose to Wall Street over 10, 20, and 30 years.
• Match Analysis: A check on whether your current contribution rate is capturing 100% of your employer's match (free money).
• Action Items: Ranked, specific steps (e.g., "Switch VFIAX for FXAIX and save $34/year").
• AI Chat: A conversational agent that holds the context of your report to answer specific follow-up questions.

How we built it

We built Fidelio from scratch in 10 hours at the Claude Builder Club @ NJIT.

• Frontend: We used Next.js 15 (App Router) with Tailwind CSS v3 and shadcn/ui to create a sleek, dark-first, highly responsive dashboard.
• AI Engine: We bypassed traditional, fragile OCR and utilized Claude Sonnet 4.6 for its native multimodal vision capabilities. Claude literally "sees" the document's spatial layout, easily associating fees with specific ticker symbols across messy tables.
• Data Orchestration (MCP): To ensure our financial advice was grounded in reality, we built a Model Context Protocol (MCP) layer. Claude dynamically calls tools to fetch live data from Finnhub (fund prices), FRED (macro risk-free rates), and SEC EDGAR (verifying expense ratios against live XBRL filings).
• Real-Time UX: We implemented Server-Sent Events (SSE) in our API routes to stream Claude's extraction, tool use, and analysis back to the client in real-time.

Challenges we ran into

1. The OCR Bottleneck: Traditional text-parsing strips away the spatial formatting of financial tables, making it impossible to connect a fund name on the left with an expense ratio on the right. We solved this by leaning entirely into Claude 4.6's spatial reasoning. 2. Hallucination Risk: When dealing with people's retirement money, the AI cannot guess. We had to heavily prompt-engineer the system and bind it to our MCP tools so it would only output verifiable data from the SEC and Finnhub. If it couldn't verify a fee, it was instructed to explicitly state that. 3. Streaming Complex States: Orchestrating a seamless UI loading state while an LLM is reading a PDF, calling three different APIs, and streaming back structured JSON chunks via SSE was a massive state-management challenge in React.

Accomplishments that we're proud of

• Zero Data Retention (ZDR): We designed the architecture to be inherently private. Documents are processed in-memory as buffers and immediately purged. Nothing touches a database.
• Live Tool Use: Getting Claude to successfully negotiate the Model Context Protocol to hit the SEC EDGAR API and normalize 20+ opaque accounting tags (like SalesRevenueNet) into a single, reliable metric.
• Speed to Value: We successfully condensed a multi-day financial auditing process down to a 30-second drag-and-drop experience, built completely within a 10-hour hackathon window.

What we learned

We learned that building trust in AI FinTech requires more than just an intelligent model; it requires provable architecture. Users will only trust an AI's financial recommendation if they understand exactly how it arrived at that number. By exposing the live MCP tool calls in the UI loading state (e.g., > CALLING TOOL: stock_compare...), we turned the AI's "thinking" process into a feature that builds user confidence.

What's next for Fidelio

We want to expand Fidelio from a diagnostic tool into an active financial advocate:

• Automated Plan Switching: Integrating with brokerage APIs to let users execute the recommended fund switches with one click.
• SECURE 2.0 Alerts: Implementing live legislative alerts (via FRED) to notify users aged 60-63 of new "super catch-up" contribution limits.
• Multi-Document Temporal Memory: Allowing users to securely save a profile so Fidelio can track their portfolio's fee reduction and health score quarter-over-quarter.

Built With

• anthropic-api
• claude-sonnet-4.6
• finnhub-api
• fred-api
• model-context-protocol
• next.js
• react
• sec-edgar-api
• server-sent-events
• shadcn-ui
• tailwind.css
• typescript