StyleSense AI

An AI shopping assistant for a fashion store. You describe what you want in plain language, like "something light for a summer wedding, under ₹5,000", and it answers with real products from the catalogue, found by semantic similarity, with the reply streaming in as it is generated.

Live demo: https://stylesense-ai-s5v8.onrender.com (free hosting, so the first load can take a minute to wake up)

The demo shows catalogue filtering, adding to the bag, and an AI stylist request. Click it for the full-quality MP4.

What it does

• Semantic product search. Product descriptions are embedded as 768-dimension vectors in Postgres with pgvector. A query is embedded the same way and matched by cosine distance, so "something floaty for warm weather" finds relevant items even when those words appear in no description.
• Streaming chat over WebSockets.
• Provider fallback: Gemini first, Groq when Gemini has a transient failure.
• Catalogue browsing with faceted filters, cursor pagination, and product pages.
• Cookie-based JWT auth with refresh-token rotation and blacklist on logout.

Architecture

                         ┌─────────────────────────────┐
   Browser ──── HTTPS ──▶│  Render (single container)  │
            └─ WSS ─────▶│                             │
                         │  Daphne (ASGI)              │
                         │   ├─ /          React SPA   │  ← same origin
                         │   ├─ /api/      DRF         │
                         │   └─ /ws/chat/  Channels    │
                         └──────────────┬──────────────┘
                                        │
                          ┌─────────────▼─────────────┐
                          │ Neon Postgres + pgvector  │
                          └───────────────────────────┘
                                        │
                          ┌─────────────▼─────────────┐
                          │  Gemini  →  Groq fallback │
                          └───────────────────────────┘

Trade-offs

Choice	Why	Cost
One container serving the SPA and API	Same origin keeps the auth cookie on SameSite=Lax. A separate frontend domain would need SameSite=None, which Safari and Brave block as a third-party cookie.	Frontend and backend deploy together.
No Redis	ChatConsumer streams only to the socket that asked and never uses groups, so the in-memory channel layer is enough.	Must change before running more than one process.
Neon instead of Render Postgres	Render's free database is deleted after 30 days. Neon's free tier doesn't expire and supports pgvector.	One more provider to manage.

Tech stack

• Backend: Django 6, DRF, Channels + Daphne, Postgres 16 + pgvector, google-genai, groq, uv
• Frontend: React 19, TypeScript, Vite, Tailwind CSS v4, TanStack Query, React Router
• Infrastructure: Docker (multi-stage), Render, Neon, GitHub Actions

Running locally

Requires Docker, uv, and Node 22 with pnpm.

git clone https://github.com/aju773/stylesense-ai.git
cd stylesense-ai
docker compose -f backend/compose.yml up -d    # Postgres + pgvector

Backend:

cd backend
cp .env.example .env    # then fill in DJANGO_SECRET_KEY and your AI keys
uv sync
uv run python manage.py migrate
uv run python manage.py seed_products
uv run python manage.py seed_embeddings   # calls the Gemini API; idempotent
uv run python manage.py runserver

Frontend:

cd frontend
pnpm install
pnpm dev

Open http://localhost:5173. Vite proxies /api and /ws to Django on :8000, the same single-origin setup as production.

Deployment

Pushing to main deploys. Render builds the Dockerfile per render.yaml, runs migrations and catalogue seeding before the switch, and only routes traffic once /healthz answers, so a container that fails to boot leaves the previous version serving.

Embeddings are seeded by hand from a local machine pointed at the production database. They call the Gemini API, and a rate limit there shouldn't fail an unrelated deploy.

CI runs lint, typecheck, the frontend build, Django's --deploy checks, a missing-migration check, and a Docker build on every push.