- Village AI uses two models of different sizes, routed by task complexity. This is not a fallback mechanism — each model is optimised for its role. +
+ Village AI uses multiple specialized models, each fine-tuned for a specific community type. The routing layer selects the appropriate model based on the tenant’s product type. All models operate under the same governance stack.
-- Handles help queries, tooltips, error explanations, short summaries, and translation. Target response time: under 5 seconds complete. +
+ Generalist model serving neighbourhood communities, governance bodies, and committees. Also serves as fallback for community types without a dedicated model.
-- Routing triggers: simple queries, known FAQ patterns, single-step tasks. +
+ Trained on te reo Māori content, whakapapa structures, and tikanga documentation. Highest indigenous domain accuracy across all variants.
- Handles life story generation, year-in-review narratives, complex summarisation, and sensitive correspondence. Target response time: under 90 seconds. +
+ Trained on Anglican parish governance, Book of Common Prayer, vestry procedures, and liturgical calendar. Serves parish and diocesan communities.
-- Routing triggers: keywords like "everything about", multi-source retrieval, grief/trauma markers. +
+ Trained on family storytelling, genealogy, heritage preservation, and inter-generational content. Highest overall FAQ accuracy. +
++ Trained on CRM, invoicing, time tracking, and professional services content. Serves business tenants and platform operations. +
++ Conservation, diaspora, clubs, and alumni models are trained when the first tenant of that type is established. Until then, the community generalist model serves.
- Both models operate under the same governance stack. Routing governance is designed; ContextPressureMonitor override capability is planned. +
+ All models are fine-tuned from the same base using QLoRA. Training data is curated per community type and never mixed across domains. A deterministic FAQ layer handles known questions without model inference. Steering vectors adjust model behaviour at inference time without modifying weights.
@@ -178,14 +200,14 @@- Each community trains a lightweight LoRA adapter on its own content — stories, documents, photos, and events that members have explicitly consented to include. This allows Village AI to answer questions like "What stories has Grandma shared?" without accessing any other community's data. +
+ Each community type (whānau, episcopal, business, family, etc.) has a dedicated fine-tuned model trained on domain-specific content. The model learns the vocabulary, governance patterns, and cultural framing appropriate to that community type. Tenant data isolation is maintained — no tenant’s content is used in another tenant’s training data.
-- Adapters are small (50–100MB). Consent is per-content-item. Content marked "only me" is never included regardless of consent. Training method: QLoRA fine-tuning with governance-validated data. +
+ Specialization is triggered when the first tenant of a new type is established. Training method: QLoRA fine-tuning with governance-validated, curated corpora.
- Honest caveat: Layer A (inherent governance via training) has been empirically validated across multiple training runs with consistent governance compliance. Layer B (active governance via Village codebase) has been operating in production for 5 months. The dual-layer thesis is demonstrating results, though evaluation remains self-reported. Independent audit is planned. + Honest caveat: Layer A (inherent governance via training) has been empirically validated across multiple training runs with consistent governance compliance. Layer B (active governance via Village codebase) has been operating in production since October 2025. The dual-layer thesis is demonstrating results, though evaluation remains self-reported. Independent audit is planned.
- Why consumer hardware? The SLL thesis is that sovereign AI training should be accessible, not reserved for organisations with data centre budgets. A single consumer GPU can fine-tune a 7B model efficiently via QLoRA. The entire training infrastructure fits on a desk. + Why consumer hardware? The SLL thesis is that sovereign AI training should be accessible, not reserved for organisations with data centre budgets. Consumer-grade GPUs can fine-tune 14B models efficiently via QLoRA. The entire inference infrastructure fits on a desk.