# LANDVEX — AI Intelligence Pipeline
**Beslutad:** 2026-06-21, Erik Svensson
**Princip:** Lokal first, Claude last. Lägsta möjliga token-kostnad.

---

## Arkitektur

```
quiXzoom bild (JPEG)
       │
       ▼
[STEG 1 — LOKAL, GRATIS]
YOLOv8n (5MB modell)
→ Object detection
→ Tags: power_pole, cable, graffiti, signage, person, vehicle, building
→ Körtid: ~80ms/bild på CPU

       │
       ▼
[STEG 2 — LOKAL, GRATIS]  
MobileNetV3 / CLIP-lite
→ Condition score 1-10
→ Vacancy: open/closed/uncertain
→ Körtid: ~50ms/bild

       │
       ▼
[STEG 3 — REGEL-BASERAT, GRATIS]
Python rule engine
→ Beräknar Infrastructure Risk Index
→ Beräknar Contradiction potential
→ Jämför mot officiell data (cache)
→ Producerar structured JSON
→ Körtid: ~5ms

       │
       ├─► 95% av bilder → direkt till databas (ingen Claude)
       │
       ▼
[STEG 4 — CLAUDE, BARA VID BEHOV]
Triggas ENDAST om:
- Contradiction potential > 0.7
- Okänd objektkategori (confidence < 0.4)
- Första bild från ny koordinat/stad

Claude-prompt: ultra-kort, strukturerad output
→ ~150 tokens in, ~100 tokens ut per bild
→ Kostnad: ~$0.0004 per triggad bild
```

---

## Structured Observation Schema

```json
{
  "observation_id": "qx-bkk-20260621-001",
  "lat": 13.7420, "lng": 100.5571,
  "city": "bangkok", "district": "sukhumvit",
  "timestamp": "2026-06-21T12:18:00Z",
  "zoomer_level": "L2",
  "image_hash": "sha256:...",

  "detected_objects": [
    {"class": "power_pole", "confidence": 0.94, "bbox": [...]},
    {"class": "cable_bundle", "confidence": 0.87, "bbox": [...]},
    {"class": "graffiti", "confidence": 0.71, "bbox": [...]},
    {"class": "commercial_sign", "confidence": 0.89, "bbox": [...]}
  ],

  "computed_scores": {
    "infrastructure_risk": 8.2,
    "condition_score": 3.1,
    "vacancy_status": "occupied",
    "graffiti_density": "low",
    "cable_disorder_severity": "critical"
  },

  "contradiction": {
    "potential": 0.84,
    "official_classification": "infrastructure_approved",
    "observed_deviation": "critical_cable_disorder",
    "claude_triggered": true,
    "narrative": "Official data: approved. Observed: critical cable bundling, fire risk."
  },

  "tags": ["electrical_hazard", "cable_disorder", "graffiti", "commercial_active"],
  "source_first": {
    "type": "observation",
    "verified_by": "ai_v1.2",
    "confidence": 0.87
  }
}
```

---

## Claude-prompt (minimalt)

Bara triggas vid contradiction_potential > 0.7:

```
SYSTEM: You are a urban intelligence classifier. Output JSON only.

USER: Location: Bangkok Sukhumvit. Objects: power_pole, cable_bundle(critical), graffiti(low). 
Official status: infrastructure_approved. 
Output: {"contradiction_score":0-1, "risk_tags":[], "narrative":"<20 words"}
```

Kostnad per prompt: ~$0.0003. Med 5% trigger-rate på 1000 bilder/dag = $0.15/dag.

---

## Implementation — server-side script

Fil: /opt/amos/api/landvex-vision/process_image.py

Kräver: ultralytics (YOLOv8), Pillow, numpy
Installeras en gång: pip install ultralytics pillow numpy anthropic

---

## Träningsstrategi

### Fas 1 — Bootstrap (nu)
- Labla 500 bilder manuellt (Erik + Johan, 2h)
- Kategorier: infrastructure_risk, vacancy, condition, graffiti
- Verktyg: Label Studio (gratis, self-hosted)

### Fas 2 — Active learning (vecka 2-4)
- Modellen flaggar bilder med låg confidence → manuell labeling
- Varje ny stad bootstrappas med 50-100 bilder

### Fas 3 — Temporal learning (månad 2+)
- Samma koordinat fotograferas igen 30 dagar senare
- Change detector: delta mellan t0 och t1
- Automatisk träningsdata utan manuellt arbete

### Fas 4 — Contradiction fine-tuning
- Alla bilder där Claude ändrade klassificeringen → träningsdata
- Målet: Claude-triggern sjunker från 5% → 1% → 0.1%