StyleSense

Inspiration

I kept watching color-analysis and Kibbe content online, then still bought clothes that looked wrong on me. StyleSense started as a personal answer: what actually flatters *me* — not another generic try-on toy. For the Backblaze Gen Media hackathon, the missing piece was durability: Runway CDN links expire, credits are expensive to re-spend, and judges care about pipeline + provenance, not just pretty frames.

What it does

Onboarding — face + body photos → color season + Kibbe profile.
Aria (agentic stylist) — LangGraph + Claude tools: URL lookup, wardrobe search, Suits/Borderline/Avoid verdicts, confirm-gated try-on & wardrobe add & save outfit.
Studio — multi-item virtual try-on + event scenes.
Genblaze → B2 — try-on stills ingested with SHA-256 manifests; image-to-video animate sinks to Backblaze B2 under per-user hierarchical keys.

How I built it

Frontend: Next.js 14 (Vercel) + Supabase Auth
Backend: FastAPI on Render · LangGraph stylist (graphs/aria_graph.py) · services/genblaze_media_service.py
Storage: Supabase (hot CDN for UI) + Backblaze B2 (durable archive)
Models: Runway gen4_image / video · Anthropic Claude · Genblaze Pipeline + RunwayProvider + ObjectStorageSink

What I learned

Agentic UX needs human-in-the-loop confirms when tools spend credits.
Gen media apps die without durable storage + manifests — B2 + Genblaze made “demo magic” into an auditable pipeline.
Production readiness is half the prize: auth, caps, health (b2_configured), and a judge-loginable seed account matter as much as the model call.

Challenges

Pooler vs direct Supabase URLs (IPv6-only db.*.supabase.co broke Render).
Keeping Genblaze/B2 visible for judges without turning the product UI into an SDK advertisement.
Credit budget discipline (turbo vs full quality; animate ~60–100 cr).