Homie.AI — MVP system architecture Diagram

Application · orchestration layer

Decides what happens next: fixed workflow with two bounded agents.
Design: room + user understanding from photo, budget, style and saved preferences.
Sourcing: choose retrieval/refinement steps; rank real catalog products and explain picks.
Each agent needs scoped tools, structured output, timeout, retry/step budget and escalation.
Intake, authorization, approval and checkout remain deterministic.

Application · core services

Performs reusable capabilities: vision/generation, style knowledge, catalog ranking, cart.
Gemini image model; Replicate-hosted ControlNet + SDXL fallback (draft).
Generation service uses model providers and saves designs in generated assets.
Both agents consume design knowledge; orchestrator supplies saved user preferences.

Data stores · persistent storage

Postgres on Supabase or Neon (draft); asset/object storage location TBD.
Session state is separate from long-term preference memory.
Memory stores preferences; agents interpret them for personalization.
Feedback updates are user-scoped, consent-aware, editable and deletable.

Cross-cutting services

Evaluation: image QA, product relevance, recommendation quality, user satisfaction.
Guardrails: content policy, schema/output validation and server-side approval enforcement.
Observability: traces, failures, latency, cost; PostHog or Mixpanel (draft).

Technical decisions & MVP assumptions

Single room; web client; hosted models; 5-week build.
Two agents + deterministic orchestration approved; vendor stack remains draft.
Checkout undecided: per-item retailer redirect or purchase within Homie.AI.
Visualization 2D vs 3D; concept-first vs real-product-grounded designs still to confirm.

Top quality goals

Helpful, realistic designs and recommendations.
Fast and smooth experience; secure and private data.
Retain historical customer preferences for personalization, with user control.

Open technical questions · engineering

Prompting only or style retrieval / RAG?
Ranking: rules, embeddings + metadata, or reranker?
Real-time furniture lookup APIs vs pre-loaded / permitted scraped catalog?
Real-time vs batch updates; price/stock freshness?
Safety, tone, room-fit uncertainty and fallback behavior?
Multi-retailer cart, checkout scope and conversion tracking?

Tech stack · initial draft

Generation: Gemini; Replicate ControlNet + SDXL fallback.
Backend: Python + FastAPI. Frontend: Next.js + component library.
Product data: Rakuten / Awin feeds into Postgres.
Hosting: Vercel + Supabase / Neon; FastAPI execution host TBD.
Analytics: PostHog or Mixpanel.