-
-
Database!
FinBytes — Hackathon Story
Category: Financial Inclusion
Problem
Financial news is written for professional, not the average person. It's jargon-heavy, long-form, and completely inaccessible to anyone who didn't study finance.
But those same people are spending hours a day scrolling TikTok. The attention is there — it's just pointed at the wrong things.
We asked one question: what if the format that makes entertainment addictive could make financial literacy addictive instead?
What It Does?
FinBytes turns high-impact financial news into swipeable micro-learning cards — think TikTok, but every scroll makes you smarter about money.
Each card presents one story four ways: a plain-English overview, an ELI5 analogy, a concrete actionable takeaway, and a social reaction. Cards are ranked by AI impact score so the most market-moving stories always surface first. Your MyDigest feed is personalised to the categories you care about from the moment you sign up.
The app tracks your daily reading streak and cards read — because habit beats motivation every time.
How We Built It?
Flutter frontend with a custom Deep Navy / Neon Green design system built.
Python pipeline in three stages: Alpha Vantage scrapes live news across seven categories → Gemini 2.5 Flash generates all four summary formats per article → an impact scorer filters out anything below 40/100 so low-significance noise never reaches the feed.
Supabase for auth and database — profiles store category preferences, streak data, and reading history per user.
Challenges We Ran Into
Inconsistent AI output. Gemini occasionally returned summaries as raw JSON arrays instead of plain text. We wrote a defensive parser in the data model that detects the format and cleans it at read time — silently, so the UI never knows.
Layout overflow hell. Category badges like "PERSONAL FINANCE" and "CRYPTO CURRENCY" kept blowing out their containers. A navigation toggle that looked fine on one device clipped on another. The fix that finally worked: stop using fixed widths entirely.
Auth timing. Supabase leaves currentUser null immediately post-signup when email confirmation is enabled. Solved by passing userId explicitly from the signup response through to onboarding, so the profile saves regardless of session state.
Accomplishments That We're Proud of
- A fully automated pipeline from raw article to four-format AI summary with zero human touch
- Personalised feeds that work from the very first login
- An impact filter that means every card in the feed has earned its place
- A polished, intentional UI that doesn't look like it was built in 24 hours
What We Learned
LLM output consistency is an engineering problem, not just a prompt problem. Defensive parsing in the client is non-negotiable.
Product thinking matters even under time pressure. The four-tab card format came from asking "what does someone need to feel genuinely informed?" — not "what's quickest to build?" That question shaped everything.
What's Next for FinBytes
Quiz cards — every fourth card becomes a comprehension question. Get it right, extend your streak. Passive scrolling becomes active learning.
Smarter personalisation — weight the feed on reading time and engagement, not just category selection.
Breaking news alerts — push notification when a story scores 85+ in your chosen categories.
Social layer — see how other readers reacted, share cards to Stories, follow friends' streaks.
Web app — the backend is already platform-agnostic. A Next.js client is the obvious next step.
Log in or sign up for Devpost to join the conversation.