ReelAlpha

Investment indicator based on finfluencers

Comment

ReelAlpha — Hackathon Submission

Inspiration

82% of young investors follow finance creators on Instagram to make stock decisions. But with no way to aggregate what multiple creators are saying about the same ticker, or track whether they've actually been right, casual investors are drowning in content yet flying blind.

What it does

ReelAlpha automatically monitors Instagram finance influencers and extracts stock sentiment from their Reels using AI. It surfaces a plain-English consensus across all creators a user follows ("5 of 6 creators are bullish on NVDA"), enriches signals with live Yahoo Finance data, and tracks every creator's accuracy over time so users know who's actually worth listening to.

How we built it

  • Frontend: Next.js 16 + React 19 + Tailwind + shadcn/ui on Vercel
  • Backend: Flask 3, event-driven via SQS + ECS Fargate worker
  • AI: Two-stage Gemini pipeline — Flash Lite for multimodal video/audio extraction, Flash for Yahoo Finance-grounded refinement — both using Structured Outputs (Pydantic schemas)
  • Data: BrightData webhooks for Reel ingestion, Apify for Yahoo Finance prices, RDS PostgreSQL + DynamoDB for storage
  • Auth: Clerk (JWT middleware on all dashboard routes)

Challenges we ran into

  • Ticker hallucination — Gemini would confidently output wrong tickers. Fixed with strict prompt constraints and a fail-silent rule: omit the signal entirely rather than guess.
  • Deduplication — BrightData can push the same Reel multiple times. Solved with a DynamoDB GSI on reel_id so duplicate webhooks are a no-op.
  • Accuracy tracking — Tying a stock price at time-of-call to an outcome weeks later required a careful stock_price_snapshots schema with a pending → win/loss lifecycle

Accomplishments that we're proud of

A fully end-to-end automated pipeline with zero manual steps: Reel posted → BrightData webhook → Lambda → SQS → ECS Worker → Gemini extracts sentiment → Yahoo Finance enriches it → result appears in the dashboard. Plus a real accuracy engine that grades every signal as a win or loss against actual historical prices — influencer tiers are earned, not cosmetic.

What we learned

Structured Outputs (Pydantic schemas via Gemini's response_schema) eliminated an entire class of JSON parsing bugs and made AI output reliable enough to build business logic on. The hardest part of an AI product is data quality, not the model.

What's next for ReelAlpha

  • Autopilot Agent — human-in-the-loop chat where users describe a thesis and the agent searches across all tracked Reels, surfaces creator signals, and challenges with counterpoints
  • Push alerts — notify users when a high-accuracy creator posts on a watchlist ticker
  • TikTok + YouTube Shorts — expand beyond Instagram
  • Stripe billing — Free / Pro / Elite tiers

Built With

Share this project:

Updates