Shaka — one score for the whole ocean.

Planning a dive day is a fragmented research loop. Before getting in the water, a spearfisher cross-checks swell, tides, visibility, wind, water temperature, and what's being caught nearby — typically across five or more sites. Even after all that, the decision is still a guess, because nothing ties the ocean physics to a specific entry point.

And the human intel is noisy: a forum post mentioning a tuna used as frozen chum can read like a catch report if you parse it naively. So the problem isn't just aggregation — it's trustworthy synthesis of satellite truth and messy human reports, with honesty about what's unknown.

Spearfishers and shore/boat divers deciding where and when to go — mostly the evening before or the morning of. It's a field tool, used on a phone at the coast, not a desktop dashboard. Power users dig into satellite layers and animated forecast maps; anglers check regional dock totals and AI-written trend summaries.

From map to decision

• Explore — a full-screen map with score-colored markers and a horizontal carousel of spots (a Surfline-style 70/30 split).
• Spot detail — three tabs: Conditions (the score breakdown, swell, tides, satellite visibility), Forecast (7-day), and Guide (marine-protected-area rules and regulations).
• Reports — regional chips across the West Coast with AI-written key insights and species catch trends (a 3-day window vs the prior 3 days).
• Charts — an animated ocean-forecast map and a NASA satellite-imagery viewer.
• Profile — saved and custom-pinned spots, unit preferences.

A Kotlin/Ktor backend on Railway ingests from many sources on staggered schedules, scores every spot, and serves a PostGIS-backed REST API. A separate Python pipeline downloads Copernicus and ECMWF fields and encodes them as map tiles served from a Cloudflare CDN. The Flutter app renders the map, the scores, and the animated layers.

The scoring engine

The 0–100 Shaka Score weights four factors, and decays its confidence honestly as the forecast horizon lengthens:

The remaining 15% is a solunar factor. When a satellite input is unavailable, the score says so rather than faking calm conditions.

• Satellite pipelines — OAuth into Copernicus with a circuit breaker for a service that dislikes concurrency; ERDDAP bbox queries with progressive expansion; a Python pipeline that turns ocean fields into WebP tiles for an animated map.
• Geospatial at the core — spots stored as PostGIS geography with radius queries for "what's near me."
• AI as a trust layer — a Groq-hosted model (with an OpenAI fallback) classifies what was actually caught and writes terse, mobile-friendly summaries; the numeric dock totals are never AI-rewritten.