Inspiration
The average financial advisor charges $200-400 per hour. Most Americans — especially students, gig workers, and first-generation immigrants — will never afford one. Yet the decisions they make about insurance, debt, and savings have lifelong consequences. We saw friends sign leases without understanding them, carry the wrong insurance, and fall into predatory loan traps simply because no one explained the fine print. SparkyFi was born from a simple belief: financial guidance shouldn't be a luxury.
What it does
SparkyFi is an AI-powered financial guide that gives everyone access to personalized financial advice: Talk to Sparky — voice or text conversations with an AI advisor that understands your actual financial situation Upload any document — bank statements, leases, insurance policies, loan agreements — get plain-english breakdowns, red flags, and action items instantly Live financial dashboard — real-time risk score, spending analysis, emergency alerts, and savings opportunities based on your transactions Insurance Intelligence — personalized insurance recommendations grounded in 2,800+ records scraped directly from State Farm and CFPB Bank connection — Plaid integration for automatic transaction analysis
How we built it
The stack was chosen for speed, accuracy, and real-world grounding: Frontend: Next.js 14 + Tailwind CSS + Framer Motion for the glassmorphism UI AI: Google Gemini 2.5 Flash for PDF vision analysis and conversational AI RAG Pipeline: Pinecone vector database with 2,800+ records scraped via BFS crawler from statefarm.com and consumerfinance.gov — every recommendation is grounded in real product data, not hallucinated Voice: Web Speech API for input, ElevenLabs for natural voice output Bank Data: Plaid sandbox for real transaction data Deployment: Vercel The RAG pipeline was the core technical investment. We built a BFS web crawler that scraped 300+ State Farm pages — every insurance product, coverage option, discount, and claims process — plus CFPB consumer education content. This means when Sparky recommends renters insurance, it's pulling actual State Farm pricing and coverage details, not making things up.
Challenges we ran into
PDF analysis latency — Gemini processing large PDFs took 30-60 seconds. We optimized by running RAG queries in parallel with PDF parsing and tightening prompts significantly. RAG dimension mismatch — Our initial Pinecone index was created with wrong dimensions for the embedding model, causing silent failures. Debugging this cost hours. Hydration errors — Next.js SSR/client mismatches kept appearing as we iterated on the UI. Solved by clearing .next cache and adding suppressHydrationWarning. Web Speech API — no-speech errors on Chrome required careful UX guidance to users to speak immediately after clicking.
Accomplishments that we're proud of
Built and deployed a fully functional AI financial advisor in one hackathon Scraped and indexed 2,800+ real records from State Farm and CFPB — no hallucinated data PDF vision analysis that actually reads leases and flags predatory clauses in real documents Voice agent with ElevenLabs that makes the experience feel like talking to a real advisor A glassmorphism UI that looks polished enough to be a real product
What we learned
RAG is only as good as the data you put in it — spending time on real data scraping paid off more than any prompt engineering Gemini's PDF vision capability is genuinely powerful for financial document analysis UX matters as much as AI — the way information is presented (red flags in a grid, risk score as a gauge) determines whether users actually act on it Building for accessibility means voice-first, not just screen-reader compatible
What's next for SparkyFi
Proactive alerts — push notifications when spending patterns change or a bill is due Multi-language support — Spanish first, to serve the largest underserved demographic State Farm agent connection — direct handoff to a local State Farm agent when the user is ready to purchase Financial goal tracking — set a savings goal, track progress, get nudges when off track
Log in or sign up for Devpost to join the conversation.