Architecture

Stack decisions, AI pipeline routing, cost model, and iOS PWA constraints.

# Stack Decisions

Layer	Choice	Rationale
Frontend	SvelteKit	Lightweight, SSR+SPA, good offline story for write-heavy PWA
Local DB	Dexie (IndexedDB)	Offline-first, sync-ready, append-heavy single-author pattern
AI Runtime	ONNX Runtime Web	On-device inference via WebAssembly, WebGPU fallback
Object Storage	Cloudflare R2	S3-compatible, zero egress fees ($150/mo vs $4,710/mo Supabase at 10K MAU)
Backend	Supabase	Postgres + PostGIS + pgvector + Auth + Realtime + Storage
Maps	MapLibre GL JS	Open-source, self-hostable, pmtiles on R2 for offline tiles

# AI Identification Pipeline

User uploads media | v Client-side EXIF extraction (exifr) --> GPS, timestamp, device | v Pre-ID quality check --> blur / dark warning | |--- Plant detected? --------> PlantNet Pro API (78K species) | |--- Bird sound? ------------> BirdNET (commercial license) | +--- General image/video ---> Gemini 2.5 Flash-Lite (~$0.001) | +-- conf >= 0.8 --> Result | +-- conf < 0.8 --> Claude Haiku 4.5 (~$0.01) | +-- conf >= 0.7 --> Result | +-- Low conf --> Expert queue Offline fallback: EfficientNet-Lite0 INT8 ONNX (<3MB, top 500 species) + Regional packs (Oaxaca/Yucatan, 10-30MB, lazy-loaded into IndexedDB)

# Cost Model

Scale	Monthly Cost	Per User
10K MAU	~$2,000	$0.20
50K MAU	~$6,000	$0.12
100K MAU	~$11,500	$0.12

10K MAU breakdown: AI inference $800, object storage $400, compute $500, monitoring and misc $300. Includes Supabase Pro, Cloudflare R2, ~500K Haiku calls, 600K PlantNet Pro calls, and Gemini Flash-Lite first-pass.

# Data Model

Core tables in the Supabase Postgres schema, partitioned monthly via pg_partman. PostGIS geography types for spatial queries, pgvector for semantic species search.

Table	Purpose	Key Fields
users	User accounts and profiles	id, display_name, role, language_pref
observations	Core observation records with location and environment	id, observer_id, observed_at, location (PostGIS), moon_phase, precipitation
species	Taxonomic reference with NOM-059 categories	id, scientific_name, common_names (jsonb), nom_059_category
media_files	Photos, audio, video with EXIF metadata	id, observation_id, media_type, r2_key, exif_data (jsonb)
identifications	AI and human ID results with confidence scores	id, observation_id, taxon_id, confidence, source, is_research_grade
expert_validations	Expert review with taxonomic badge verification	id, identification_id, validator_id, verdict, notes