Intro & scope
Assistants for private individuals and households: money and paperwork (MONEY), health (HEALTH), daily life and media (LIFE) and privacy/self-hosting (PRIVACY). Users are consumers, not employers or professionals; B2B variants live in other families (DOC, HLTH, HR, LEGAL). PERS-005 is the hub: a private local stack that most other PERS cases can run on.
Scope boundaries: - PERS-001 (budget/spending) vs PERS-013 (annual tax return): 001 is continuous bookkeeping, 013 is the yearly filing; both reuse the same receipt pipeline. - PERS-008 (understand/answer any official letter) vs PERS-015 (assess and pursue a consumer claim) vs PERS-016 (GDPR requests): 008 explains, 015 judges merit, 016 exercises data rights. - PERS-002 (wellness coaching from wearables) vs PERS-021 (cycle/fertility prediction, medical-device territory) vs PERS-014 (explaining existing medical records). - PERS-006 (photo library) vs PERS-017 (wearable lifelog capture): 006 organises existing media, 017 creates new continuous captures. - PERS-007 (assistive vision) vs PERS-018 (DIY repair vision): accessibility versus task guidance.
Table
| ID | Title | Pattern | Maturity | Path | HW | AI Act |
|---|---|---|---|---|---|---|
| PERS-001 | Personal finance & budgeting assistant | MONEY | mainstream | A | HW1 | minimal |
| PERS-002 | Personal health, fitness & sleep coach | HEALTH | early-adopter | A | HW2 | depends |
| PERS-003 | Travel planning & itinerary assistant | LIFE | mainstream | A | HW2 | minimal |
| PERS-004 | Recipe, meal planning & nutrition | LIFE | mainstream | A | HW1 | minimal |
| PERS-005 | Private local AI assistant (home server) | PRIVACY | early-adopter | C | HW2 | minimal |
| PERS-006 | Photo & video library organisation | LIFE | mainstream | A | HW1 | minimal |
| PERS-007 | Assistive AI for blind & low-vision people | LIFE | early-adopter | A | HW1 | limited |
| PERS-008 | Bureaucracy helper (official letters) | MONEY | early-adopter | B | HW2 | minimal |
| PERS-009 | Family & household organisation | LIFE | early-adopter | A | HW1 | minimal |
| PERS-010 | Genealogy & family history | LIFE | early-adopter | A | HW2 | minimal |
| PERS-011 | Shopping, price comparison & deal agent | MONEY | early-adopter | A | HW2 | minimal |
| PERS-012 | Job application assistant | MONEY | mainstream | A | HW2 | minimal |
| PERS-013 | Personal tax return assistant | MONEY | mainstream | A | HW2 | minimal |
| PERS-014 | Patient-side medical report/ePA explainer | HEALTH | early-adopter | A | HW2 | depends |
| PERS-015 | Consumer-rights & legal self-help | MONEY | early-adopter | A | HW2 | minimal |
| PERS-016 | Personal data-rights agent (GDPR) | PRIVACY | early-adopter | A | HW1 | minimal |
| PERS-017 | Wearable lifelogging & memory assistant | LIFE | early-adopter | A | HW1 | limited |
| PERS-018 | Visual DIY repair assistant | LIFE | early-adopter | A | HW2 | minimal |
| PERS-019 | Child online safety & parental assistance | LIFE | mainstream | A | HW1 | limited |
| PERS-021 | Cycle tracking & fertility prediction | HEALTH | mainstream | A | HW1 | high |
(Path: A = SaaS/API with EU hosting, B = hybrid, C = self-hosted.)
Shared stack
Common building blocks: LLM + embeddings + vector DB (retrieval over personal documents), OCR/doc-parse, VLM, ASR/TTS, a workflow/agent layer and guardrails. One personal knowledge store serves all PERS cases; sensitive health and child data stay in separate encrypted collections.
HW1 (mini-PC / laptop NPU, 16 GB RAM, no dedicated GPU) - LLM: 3-8B quantised small model (e.g. Qwen/Llama/Mistral class) via Ollama or llama.cpp; embeddings: bge-m3 or multilingual-e5-small. - Vector DB: SQLite-vec or Chroma/Qdrant single node. OCR: Docling/Tesseract/PaddleOCR. ASR: whisper.cpp (small/medium). TTS: Piper. - Vision: small VLM (Qwen-VL / SmolVLM class), CLIP/SigLIP for photo search; classifiers via ONNX Runtime. Time-series/tabular: scikit-learn, LightGBM. - Automation: n8n or Home Assistant; guardrails: rule filters + small classifier. - Fits: PERS-001, 004, 006, 007, 009, 016, 017, 019, 021.
HW2 (workstation/home server, 24-48 GB VRAM GPU, 64 GB RAM) - Everything in HW1 plus: 14-32B model or MoE with vLLM/Ollama (reasoning model for tax, legal, letters), larger VLM, whisper-large. - Serving: Open WebUI + LiteLLM gateway; RAG with Qdrant/pgvector + reranker; agent framework (LangGraph or similar) with browser automation (Playwright) for PERS-003/011/016. - Fits: PERS-002, 003, 005, 008, 010, 011, 012, 013, 014, 015, 018.
Cloud fallback (Path A/B): EU-region API (zero data retention, DPA, no training on inputs) behind the same LiteLLM gateway; route by sensitivity: health, children, tax and ID documents local-first; travel, shopping, recipes may use cloud.
Cross-family reuse: document pipeline shared with DOC (DOC-002/014/017/018); speech with SPEECH-001; vision with VIS-025/030; agents with AGENT-001/003; knowledge store with KNOW-002/016/018; health data model with HLTH-011.
Governance defaults: encrypt at rest, local keys, per-collection access, audit log of agent actions, human confirmation before submit/pay/send; PERS-021 requires medical-device (MDR) conformity and is AI Act high-risk; PERS-002/014 depend on intended purpose (wellness vs medical).