Appendix C

Technology Stack Reference

← Back to companion repository

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.

ProviderModelContextInput £/1MOutput £/1MAWB Use
GoogleGemini 3.5 Flash1M£1.18£7.09✓ Primary
GoogleGemini 3.1 Pro1M£1.57£9.45✓ Complex
OpenAIGPT-5.51M£3.94£23.62✓ Alternative
OpenAIGPT-5 Mini1M£0.20£1.57✓ Budget
AnthropicClaude Sonnet 4.6200K£2.36£11.81✓ Reasoning
AnthropicClaude Opus 4.81M£3.94£19.69⚠ Evaluate
MetaLlama 4 Maverick128KInfra 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

LibraryVersionPurposeChapter
Python3.11+Primary languageAll
FastAPI0.110+API frameworkAll services
Pydantic2.6+Data validation & settingsAll
LangGraph0.1+Agent state machinesCh03
ChromaDB0.4+Vector databaseCh04, Ch12
RAGAS0.1+RAG evaluationCh04
XGBoost2.0+Credit & fraud modelsCh06, Ch08
SHAP0.44+Model explainabilityCh06, Ch07, Ch08
MLflow2.11+Experiment trackingCh14
DVC3.40+Data versioningCh14
Airflow2.9+Pipeline orchestrationCh14, Ch15
rank-bm250.2+BM25 sparse retrievalCh04
adjustText1.0+Chart label placementAll figures
pytest7.4+Testing frameworkAll

C.3 — Cloud Services (AWS eu-west-2, UK)

ServiceAWB UseData ResidencyPRA/DORA Notes
ECS FargateAll 23 AI service containerseu-west-2 (UK)ICT asset registration required
RDS PostgreSQLAudit logs, model outputs, COREP dataeu-west-2 (UK)7-year retention policy
ElastiCache RedisAgent memory, API cachingeu-west-2 (UK)90-day TTL for working memory
S3Data lake raw/curated zones, model artefactseu-west-2 (UK)7-year retention for raw zone
MSK (Kafka)Real-time fraud feature streamingeu-west-2 (UK)DORA ICT asset registration
CloudWatchMetrics, alarms, log aggregationeu-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 GroupAPI Cost/MonthNotes
Chapters 1–5 (Foundations)£0Free tier: Gemini 3.1 Pro / Flash free tier covers all examples
Chapters 6–10 (Risk domains)£5–£20LLM calls + modest PostgreSQL/Redis for local dev
Chapters 11–12 (Compliance)£5–£15AML/KYC API calls, sanctions data feed (test)
Chapters 13–15 (Enterprise)£20–£50AWS ECS + RDS + ElastiCache for integration tests
Chapter 16 (Integrated)£30–£60Full platform stack, all 23 services running