Files

T

Bernt df5e2b3e78 docs: LandveX Intelligence Lab v1.2 — MVP + Phases + Data Quality + Decision Analytics

- MVP Milestone: 'First Verified Decision'
  - Developer films with quiXzoom, imports to Lab, corrects AI,
    creates Decision Case, follows chain with full traceability
  - When this works = first complete verifiable Control Intelligence pipeline

- Three development phases:
  Phase 1 (Essential): Ingestion, Dataset Explorer, Annotation, Decision Case Viewer
  Phase 2 (Scale): Replay, Benchmark, Evaluation
  Phase 3 (Advanced): GPU Jobs, Hyperparameter Runs, Model Promotion, Canary

- Product Architecture: quiXzoom → Observations → Intelligence Lab →
  Improved Models → LandveX → Better Decisions → Feedback → Intelligence Lab
  - Two products: quiXzoom (observations), LandveX (decisions)
  - Intelligence Lab = the factory that improves both

- New areas:
  - Data Quality: Healthy/Blurred/Duplicate/Wrong GPS/Night/Rain/Occluded
    + Coverage (Roads, Buildings, Signs, Drainage, Vegetation)
  - Decision Analytics: Acceptance Rate, Ignore Rate, Accuracy,
    Insufficient Evidence, Data Collection Value
    - Business value metrics, not traditional AI metrics

Rationale: Build MVP first, prove first real workflow, then scale.
Decision Cases are the heart. Data Quality explains model performance.
Decision Analytics measure business value.

2026-07-02 12:27:16 +00:00

11 KiB

Raw Blame History

LANDVEX INTELLIGENCE LAB

Internal Development Environment for Control Intelligence


Version	1.0
Status	SPECIFICATION
Purpose	Build, test, and verify LandveX Control Intelligence before production

Core Principles

LandveX Intelligence Lab does not produce AI models. It produces verified Control Intelligence.

This is an internal tool. Never a customer product.

All artifacts are immutable and versioned.

Raw images are never modified
Annotations are versioned
Models are versioned
Evaluation reports are versioned
Decision Cases are versioned
Replay results are saved as new artifacts

Every change must be traceable to a specific experiment, model version, dataset, and decision.

Product Architecture

quiXzoom
    ↓
Observations
    ↓
LandveX Intelligence Lab
    ↓
Improved Models
    ↓
LandveX
    ↓
Better Decisions
    ↓
Feedback
    ↓
Intelligence Lab

Two products:

quiXzoom — produces observations
LandveX — produces decisions
Intelligence Lab — the factory that improves both

Repository

landvex-intelligence-lab

Separate from:

quixzoom-app
landvex-web
aamos-core

Architecture

LandveX Intelligence Lab
│
┌─────────────────┼─────────────────┐
│                 │                 │
Developer UI    Training API    Experiment API
│                 │                 │
└─────────────────┴─────────────────┘
│
Event Bus / Queue
│
─────────────────────────────────────────────────────
Storage Layer
Images │ Videos │ Missions │ Models │ Logs
S3/R2 Buckets + PostgreSQL + Neo4j
─────────────────────────────────────────────────────
│
Processing Pipelines

Dashboard
├── Models
├── Datasets
├── Annotations
├── Training
├── Evaluation
├── Decision Cases
├── Replay
├── Validation
├── Promote Model
└── Settings

Dashboard

Shows AI status, not business data.

┌─────────────────────────────────────┐
│ Datasets                            │
│ Healthy: 12                         │
│ Needs Review: 3                     │
│ Corrupted: 0                        │
├─────────────────────────────────────┤
│ Training Jobs                       │
│ Running: 2                          │
│ Queued: 1                           │
│ Completed: 47                       │
│ Failed: 0                           │
├─────────────────────────────────────┤
│ Decision Cases                      │
│ Validated: 23                       │
│ Pending: 5                          │
│ Rejected: 2                         │
├─────────────────────────────────────┤
│ Replay Jobs                         │
│ Ready: 8                            │
│ Running: 1                          │
│ Finished: 34                        │
└─────────────────────────────────────┘

Mission Replay

Click through the entire chain:

Video → Frame → Bounding boxes → Detected objects → Evidence → Finding → Decision → Business Impact

Annotation

┌─────────┬─────────────┬──────────────┐
│ Video   │ AI Suggestion│ Manual       │
│         │              │ Correction   │
├─────────┼─────────────┼──────────────┤
│         │ Object:      │ Correct?     │
│         │ Road Crack   │ YES / NO     │
│         │ Confidence:  │              │
│         │ 82%          │ Severity:    │
│         │              │ Low / Medium │
│         │              │ / High       │
└─────────┴─────────────┴──────────────┘

Decision Cases

The most important asset.

Not images. Not videos. Not AI models.

But:

Observation → Evidence → Finding → Decision → Outcome → Learning

After a few years, hundreds of thousands of verified Decision Cases. Not just a training dataset — a library of how real observations lead to real decisions and real outcomes.

Case #4232
├── Reality
├── Observation
├── Evidence
├── Finding
├── Decision
├── Outcome
└── Learning

All cases playable.

Benchmark

Compare models:

Model	Precision	Recall	F1	Latency	Decision Accuracy
YOLO v8	0.89	0.87	0.88	45ms	—
Grounding DINO	0.91	0.85	0.88	120ms	—
SAM	0.88	0.90	0.89	200ms	—
Custom	0.92	0.91	0.915	60ms	0.87

Replay

Find regressions:

Mission 213
├── Play
├── Show AI
├── Show Human Annotation
├── Differences
├── New Model
└── Old Model

Validation

Field Trials
Scenario Tests
Decision Tests
Evidence Tests
Golden Failures
Regression Tests

Promote Model (Not Deploy)

Development → Validation → Pilot → Production

Not "Deploy". "Promote Model".

Experiments

Experiments
├── EP-1.0
├── DS-001
├── DS-002
├── DS-003
├── Field Trials
└── Metrics

Link experiment protocol to real development and validation data.

Processing Pipelines

1. Ingestion Pipeline

Input: Images, video, GPS, EXIF, metadata Checks: Checksums, versioning Output: Raw Dataset

2. Dataset Pipeline

Input: Raw Dataset Checks: Sort, deduplicate, quality control, resolution, blur detection, GPS validation Output: Validated Dataset

3. Annotation Pipeline

Input: Validated Dataset Process: AI suggestions, manual correction, label versions, consensus Output: Verified Dataset

4. Training Pipeline

Input: Verified Dataset Process: Start training, hyperparameters, checkpoints, GPU jobs Output: Model Artifact

5. Evaluation Pipeline

Input: Model Artifact Metrics: Precision, recall, F1, decision accuracy, regression Output: Evaluation Report

6. Replay Pipeline

Input: Old missions, Model v14, Model v15 Process: Run both models, compare differences Output: Regression Report

7. Decision Validation Pipeline

Input: Observation Process: Full chain — Observation → Evidence → Finding → Decision → Business Impact Output: Decision Validation Report

Bucket Structure

raw-images/
raw-video/
missions/
datasets/
annotations/
training/
models/
replays/
evaluation/
decision-cases/
field-trials/
exports/
archives/

All content is versioned:

model-v14/
model-v15/
model-v16/

AI Job Queue

All jobs are asynchronous:

Upload → Queue → Worker → GPU → Storage → Notification

Not synchronous API calls.

What This Tool Collects

Model training
Annotation
Datasets
Replay
Decision chains
Validation
Regression tests
Experiments
Model promotion

New Developer Experience

A new AI engineer should open the repo and within minutes understand:

"This is the tool where we build, test, and verify LandveX Control Intelligence before anything reaches production."

Relationship to Principles

All development in Git
All experiments reproducible
All models traceable from training to validation to production
Version control and traceability

MVP Milestone: "First Verified Decision"

Definition: A developer can:

Film a real object with quiXzoom
Import material to Intelligence Lab
Review and correct AI interpretation
Create a Decision Case
Follow the entire chain from observation to decision with full traceability

When this works, you have the first complete, verifiable Control Intelligence pipeline.

Development Phases

Phase 1 — Essential (for pilot missions)

Feature	Purpose
Ingestion	Upload images/video, show metadata (GPS, time, device), store raw data immutably
Dataset Explorer	Browse missions, filter, search, open a mission
Annotation	AI suggestions, manual correction, version history
Decision Case Viewer	Observation → Evidence → Finding → Decision → Business Impact → Learning

This is the heart.

Phase 2 — Scale (when running many missions)

Feature	Purpose
Replay	Compare model versions on same mission
Benchmark	Model comparisons
Evaluation	Regression tests

Phase 3 — Advanced (when having multiple models)

Feature	Purpose
GPU Jobs	Training queue
Hyperparameter Runs	Automated experiments
Model Promotion	Development → Validation → Pilot → Production
Canary Releases	Gradual rollout

New Areas

Data Quality

Before training anything:

Images
├── Healthy
├── Blurred
├── Duplicate
├── Wrong GPS
├── Night
├── Rain
└── Occluded

Coverage:

Roads, Buildings, Signs, Drainage, Vegetation

Quickly shows why a model performs well or poorly.

Decision Analytics

Not AI metrics. Decision metrics.

Metric	Question
Acceptance Rate	How many recommendations were accepted?
Ignore Rate	How many were ignored?
Accuracy	How many proved correct?
Insufficient Evidence	How many lacked enough evidence?
Data Collection Value	How often did "Collect more data" lead to a better decision?

Closer to LandveX business value than traditional AI metrics.

ändringshistoria

Version	Datum	Beskrivning
1.0	2026-07-02	Initial specification for LandveX Intelligence Lab
1.1	2026-07-02	Added architecture, pipelines, immutability
1.2	2026-07-02	Added MVP milestone, development phases, Data Quality, Decision Analytics

STATUS

SPECIFICATION — Awaiting development decision

11 KiB Raw Blame History