Creator Spark

Inspiration

Creators told us two clear things: they want to see how earnings are calculated, and they need simple protections against fraud. Creator Spark brings those into one clear, familiar dashboard.

What it does

Fair bonuses. Every video gets two signals:
- Engagement grade: a score (0–1) mapped to a letter grade (A, B, …).
- Originality: 1 (original) or 0 (copied/near-duplicate).
Bonus rules. Only Grade B or better earns a bonus.
- Grade A → +0.20× base pay
- Grade B → +0.10× base pay
- Original videos get the full bonus; non-original get half the bonus.
- Example: base payout $100, Grade A, original ⇒ $100 + $20 = $120.
Fraud protection. Continuous checks for suspicious user agents, referrer tricks, burst spikes, bulk account creation, shared devices, and unusual gifting. If flagged, payouts freeze and the creator gets a clear report with an appeal path.
Creator dashboard.
- Overview: total earnings, followers, views, weekly earnings chart, active fraud alerts.
- Posts: per-video views/likes/comments/saves/shares, grades, and an earnings breakdown. Click into a post to see the score details, bonus math, and any fraud alerts.

How we built it (simple)

Originality scoring: sample frames from each video, create compact image hashes (“fingerprints”), store them, and compare new videos against the library to estimate similarity.
Fraud checks: look for fast, unnatural patterns and metadata mismatches; freeze and flag when thresholds are crossed.
Dashboard: mirrors a TikTok-style profile, but adds transparent scoring and payouts.

Tech we used (selected)

FastAPI + Uvicorn + Pydantic: clean, fast API with input validation.
yt-dlp, OpenCV, Pillow, ImageHash, NumPy: download videos, sample frames, make perceptual hashes, and compare fingerprints.
SQLite: lightweight store for video fingerprints.
Flask + Supabase: simple fraud-alert API and storage (timestamps, alert states, joins to posts/profiles).
Logging & Datetime: clear audit trail and alert timelines.

Challenges we ran into

Setting bonus thresholds that feel fair across niches and lengths.
Keeping originality strict enough to matter without blocking trends/remixes.
Spotting fake spikes without punishing real virality.
Explaining the why behind earnings in language creators actually read.

Accomplishments we’re proud of

A working flow end-to-end: upload → score → bonus → earnings view → fraud alert & appeal.
A clear, example-driven bonus breakdown that matches backend math.
Originality analyzer with nearest-match hints to help creators understand scores.
Fraud alerts that freeze payouts and produce a plain-English report.

What we learned

Transparency builds trust. Short, concrete explanations beat long rules.
A few strong signals (watch time, saves, real comments) carry more weight than huge metric lists.
“Fair and firm” fraud rules work best when appeals are easy and specific.

What’s next for Creator Spark

Creator app tips: “before you post” hints using the same rules.
Appeals UX: in-app resolution with clearer evidence views.
More platforms: carry your grade/bonus history across apps.
Independent review: external checks on bias and thresholds.
Education hub: simple guides on improving engagement quality and originality.

Appendix (for judges/devs)

Originality detection (how it works in brief)

Download video (URL) → sample frames → convert to images.
Compute perceptual hashes per frame (pHash/dHash/aHash).
Aggregate into a single fingerprint vector and store in SQLite.
For new videos, build the fingerprint, compare to stored vectors, and return an originality score plus nearest matches.

Key libraries: FastAPI, Uvicorn, Pydantic, yt-dlp, OpenCV, Pillow, ImageHash, NumPy, sqlite3.

Fraud detection (signals & endpoints)

Signals: odd user agents, referrer manipulation, burst spikes, bulk account creation, shared devices, suspicious gifting.
Actions: generate alert → freeze payout → notify creator → allow appeal → resolve.

Simple API (prototype):