Bernt
9a817ac82a
- Added architecture diagram: - Developer UI / Training API / Experiment API - Event Bus / Queue - Storage Layer (S3/R2 + PostgreSQL + Neo4j) - Processing Pipelines - Seven separate pipelines (not one big AI loop): 1. Ingestion — raw images/video/GPS/EXIF → Raw Dataset 2. Dataset — sort, deduplicate, quality control → Validated Dataset 3. Annotation — AI suggestions, manual correction → Verified Dataset 4. Training — hyperparameters, checkpoints, GPU → Model Artifact 5. Evaluation — precision, recall, F1, decision accuracy → Report 6. Replay — compare model versions → Regression Report 7. Decision Validation — full chain → Validation Report - Bucket structure: raw-images, raw-video, missions, datasets, annotations, training, models, replays, evaluation, decision-cases, field-trials, exports, archives - AI Job Queue: all jobs asynchronous (Upload → Queue → Worker → GPU → Storage → Notification) - Core principles updated: - All artifacts immutable and versioned - Every change traceable to experiment, model, dataset, decision - Dashboard shows: Datasets (Healthy/Needs Review/Corrupted), Training Jobs, Decision Cases, Replay Jobs - Decision Cases emphasized as most important asset Rationale: Separate pipelines make system easier to debug, improve, and swap components. Immutability aligns with E-001 Git/traceability principles. Decision Cases become the unique asset over time.
nordea-kyc-fraga3
@ f831552401
