Intro and scope
Systems that decide what to show, offer or say to whom. Two patterns: - REC.RECOMMEND: ranking items or people for a user (products, content, jobs, courses, matches, size, advice, B2B configurations). - REC.MARKETING: audience, campaign, budget and channel decisions (segmentation, next-best-action, email, ads, MMM, loyalty, retail media, influencers, GEO).
Scope boundaries: - REC-001 is the generic item recommender baseline; REC-002 (editorial content, pluralism), REC-012 (size/fit), REC-010 (B2B, rules-constrained) and REC-018 (degree rules) specialise it. - REC-008 (job seekers, recommends vacancies) is not candidate scoring for employers (that is high risk, see PRED/CONV-019); REC-016 requires mutual acceptance of a match. - REC-003 chooses the action/offer per customer; REC-006 executes email/push content and timing; REC-009 is the same logic limited to loyalty members; REC-004 provides segments to all three. - REC-007 optimises own ads and bidding; REC-013 is the retailer-side sponsored-product marketplace; REC-015 selects creators; REC-005 measures aggregate channel effect; REC-011 is visibility in AI assistants (no personal data). - REC-017 is regulated advice (MiFID II suitability), not marketing. - Search and RAG retrieval belong to KNOW; forecasting/propensity models to PRED; solvers to OPT. - AI Act: mostly minimal; limited (Art. 50) where generated content or bots reach consumers (REC-006, -007); depends/high where outputs steer employment (REC-008), education (REC-018), financial services (REC-017) or exploit vulnerabilities (REC-003, -016). GDPR: profiling needs legal basis, consent for tracking (ePrivacy), opt-out and Art. 22 safeguards; DSA Art. 27/38 recommender transparency and non-profiling option for platforms.
Overview
| ID | Title | Pattern | Maturity | Path | HW tier | AI Act risk |
|---|---|---|---|---|---|---|
| REC-001 | Product recommendations in e-commerce | REC.RECOMMEND | mainstream | A | HW2 | minimal |
| REC-002 | Content recommendations (media, news, streaming) | REC.RECOMMEND | mainstream | B | HW3 | minimal |
| REC-003 | Next-best-action & personalised marketing | REC.MARKETING | mainstream | A | HW1 | depends |
| REC-004 | Customer segmentation | REC.MARKETING | mainstream | A | HW1 | minimal |
| REC-005 | Marketing mix modelling & attribution | REC.MARKETING | early-adopter | B | HW1 | minimal |
| REC-006 | Email & campaign personalisation, send-time optimisation | REC.MARKETING | mainstream | A | HW2 | limited |
| REC-007 | Ad creative optimisation & programmatic bidding | REC.MARKETING | mainstream | A | HW2 | limited |
| REC-008 | Job & career recommendations for job seekers | REC.RECOMMEND | mainstream | B | HW2 | depends |
| REC-009 | Loyalty programme personalisation & offers | REC.MARKETING | mainstream | A | HW1 | minimal |
| REC-010 | B2B cross-sell, upsell & product configuration | REC.RECOMMEND | mainstream | A | HW1 | minimal |
| REC-011 | Generative engine optimisation | REC.MARKETING | early-adopter | A | HW0 | minimal |
| REC-012 | Size & fit recommendation | REC.RECOMMEND | mainstream | A | HW1 | minimal |
| REC-013 | Retail media sponsored-product ranking | REC.MARKETING | mainstream | A | HW2 | minimal |
| REC-015 | Influencer discovery & authenticity checks | REC.MARKETING | mainstream | A | HW2 | minimal |
| REC-016 | Reciprocal matchmaking | REC.RECOMMEND | mainstream | A | HW2 | depends |
| REC-017 | Robo-advisory & suitability-based advice | REC.RECOMMEND | mainstream | A | HW1 | depends |
| REC-018 | Course selection & degree planning | REC.RECOMMEND | early-adopter | A | HW2 | depends |
Shared stack
Common building blocks: event/interaction log and feature store (Postgres/ClickHouse, Feast), consent and preference service (CMP, opt-out, Art. 22 review), experiment and bandit layer (A/B, interleaving, holdouts), catalog/profile embeddings, rules/guardrail layer (eligibility, frequency caps, fairness), offline-to-online evaluation with logging.
HW0 (no GPU, SaaS/API) for REC-011 and the LLM parts of any REC case: - Hosted EU-region LLM API with prompt/response logging; scheduler plus Postgres; no personal data in prompts.
HW1 (CPU server or small GPU, <=8 GB) for REC-003, -004, -005, -009, -010, -012, -017: - LightGBM/XGBoost/CatBoost, implicit/LightFM, PyMC-Marketing or Meridian (MMM), scikit-learn clustering, OR-Tools/cvxpy for portfolios and slots. - Postgres + pgvector, Kafka or Redpanda for streams, Feast feature store, MLflow; optional 3-8B INT4 LLM via Ollama for segment descriptions. - Shares the feature store and experiment layer with PRED and OPT.
HW2 (single GPU, 16-24 GB) for REC-001, -006, -007, -008, -013, -015, -016, -018: - Two-tower/sequence models (PyTorch, Merlin/TorchRec), bge-m3 or multilingual-e5 embeddings, bge-reranker, Qdrant or pgvector/OpenSearch for retrieval. - vLLM/Ollama serving 7-14B instruct (explanations, subject lines, copy) plus VLM (Qwen-VL class) for creative and creator content; Langfuse tracing and guardrails. - Shares embedding service, vector DB and reranker with KNOW-002 and CONV RAG stacks.
HW3 (multi-GPU / EU private cloud, 2-8 x 40-80 GB) for REC-002: - Large-scale retrieval + ranking cascade (ANN, DLRM/transformer rankers, RL/bandit serving), streaming feature pipeline (Flink), multimodal embeddings for video/audio/text, LLM for editorial tagging and summaries; pluralism and diversity metrics, DSA transparency reporting.