The approved technology stack for the AWB-AI-2025 programme as of June 2026. All versions, costs, and configuration choices reflect AWB's production environment.
Verify current pricing before procurement — model availability and pricing change frequently.
C.1 — Approved LLM Reference (Table 1.9, June 2026)
The only LLMs referenced throughout the book. Selected on context window, structured-output reliability, £ cost at AWB's processing volumes, data residency, and DORA concentration risk. Costs converted at £1 = $1.27 (June 2026) — reference tables show standard list pricing; AWB's actual ROI figures use its effective blended rate after batch and prompt-cache discounts. Never use a model outside this list in production.
| Provider | Model | Context | Input £/1M | Output £/1M | AWB Use |
|---|---|---|---|---|---|
| Gemini 3.5 Flash | 1M | £1.18 | £7.09 | ✓ Primary | |
| Gemini 3.1 Pro | 1M | £1.57 | £9.45 | ✓ Complex | |
| OpenAI | GPT-5.5 | 1M | £3.94 | £23.62 | ✓ Alternative |
| OpenAI | GPT-5 Mini | 1M | £0.20 | £1.57 | ✓ Budget |
| Anthropic | Claude Sonnet 4.6 | 200K | £2.36 | £11.81 | ✓ Reasoning |
| Anthropic | Claude Opus 4.8 | 1M | £3.94 | £19.69 | ⚠ Evaluate |
| Meta | Llama 4 Maverick | 128K | Infra cost | — | ✗ SS1/23 burden |
Multi-LLM concentration split (DORA Art. 28, cap 70% per provider): Gemini 3.5 Flash 68% | Claude Opus 4.8 17% | GPT-5.5 15%, per book Table 1.9. Note this differs from the per-chapter agentic pipeline allocation elsewhere in the repo, which routes the 17% Anthropic share to Claude Sonnet 4.6 (Agent 5 / HITL) rather than Opus — worth reconciling before print.
C.2 — Core Python Libraries
| Library | Version | Purpose | Chapter |
|---|---|---|---|
| Python | 3.11+ | Primary language | All |
| FastAPI | 0.110+ | API framework | All services |
| Pydantic | 2.6+ | Data validation & settings | All |
| LangGraph | 0.1+ | Agent state machines | Ch03 |
| ChromaDB | 0.4+ | Vector database | Ch04, Ch12 |
| RAGAS | 0.1+ | RAG evaluation | Ch04 |
| XGBoost | 2.0+ | Credit & fraud models | Ch06, Ch08 |
| SHAP | 0.44+ | Model explainability | Ch06, Ch07, Ch08 |
| MLflow | 2.11+ | Experiment tracking | Ch14 |
| DVC | 3.40+ | Data versioning | Ch14 |
| Airflow | 2.9+ | Pipeline orchestration | Ch14, Ch15 |
| rank-bm25 | 0.2+ | BM25 sparse retrieval | Ch04 |
| adjustText | 1.0+ | Chart label placement | All figures |
| pytest | 7.4+ | Testing framework | All |
C.3 — Cloud Services (AWS eu-west-2, UK)
| Service | AWB Use | Data Residency | PRA/DORA Notes |
|---|---|---|---|
| ECS Fargate | All 23 AI service containers | eu-west-2 (UK) | ICT asset registration required |
| RDS PostgreSQL | Audit logs, model outputs, COREP data | eu-west-2 (UK) | 7-year retention policy |
| ElastiCache Redis | Agent memory, API caching | eu-west-2 (UK) | 90-day TTL for working memory |
| S3 | Data lake raw/curated zones, model artefacts | eu-west-2 (UK) | 7-year retention for raw zone |
| MSK (Kafka) | Real-time fraud feature streaming | eu-west-2 (UK) | DORA ICT asset registration |
| CloudWatch | Metrics, alarms, log aggregation | eu-west-2 (UK) | DORA incident detection |
C.4 — Developer Cost Guide by Chapter Group
All code examples can be run locally with docker-compose and the Gemini 3.5 Flash free tier. AWS services are optional for local development; mocked equivalents are provided in the test suite.
| Chapter Group | API Cost/Month | Notes |
|---|---|---|
| Chapters 1–5 (Foundations) | £0 | Free tier: Gemini 3.1 Pro / Flash free tier covers all examples |
| Chapters 6–10 (Risk domains) | £5–£20 | LLM calls + modest PostgreSQL/Redis for local dev |
| Chapters 11–12 (Compliance) | £5–£15 | AML/KYC API calls, sanctions data feed (test) |
| Chapters 13–15 (Enterprise) | £20–£50 | AWS ECS + RDS + ElastiCache for integration tests |
| Chapter 16 (Integrated) | £30–£60 | Full platform stack, all 23 services running |