Prisme

Catalogue quality audit platform for FMCG / retail. Prisme ingests 1 000 Open Food Facts products, runs a dual-branch audit pipeline (text metadata + visual assets), exposes a REST API, and serves a Next.js dashboard with CLIP-powered visual search and daily Gemini AI reports.

Live demo: prisme.stephanewamba.com

---

Architecture

graph TD
    OFF[Open Food Facts CSV] --> INGEST[ingest_off.py]
    INGEST --> BQ_RAW[(BigQuery\nproducts_selected)]

    BQ_RAW --> TEXT[Text branch\ntext_profiler + anomaly_detector]
    BQ_RAW --> VISUAL[Visual branch\ndownloader + encoder + thumbnailer + vision]

    TEXT --> SCORER[scorer.py\nfusion 60% text + 40% visual]
    VISUAL --> SCORER

    SCORER --> BQ_SCORES[(BigQuery\nproduct_scores)]
    TEXT --> BQ_ANOMALIES[(BigQuery\ntext_anomalies)]
    VISUAL --> BQ_CLIPS[(BigQuery\nclip_embeddings)]
    VISUAL --> GCS[(GCS\nprisme-assets)]

    BQ_SCORES --> API[FastAPI\nCloud Run]
    BQ_ANOMALIES --> API
    BQ_CLIPS --> API
    GCS --> FRONTEND[Next.js\nVercel]
    API --> FRONTEND

    GEMINI[Gemini 2.5 Flash] --> REPORT[report_generator.py]
    BQ_SCORES --> REPORT
    BQ_ANOMALIES --> REPORT
    REPORT --> BQ_REPORTS[(BigQuery\nreports)]
    BQ_REPORTS --> API

Loading

---

Pipeline

flowchart LR
    subgraph Text branch
        T1[text_profiler\nmetadata completeness] --> T2[anomaly_detector\nCRITICAL / HIGH / MEDIUM]
    end

    subgraph Visual branch
        V1[downloader\nfetch image_url] --> V2[encoder\nCLIP embeddings to BQ]
        V1 --> V3[visual_scorer\nresolution + sharpness]
        V1 --> V4[thumbnailer\n128 / 256 / 512 px to GCS]
        V1 --> V5[vision.py\nVision API centration]
    end

    subgraph Fusion
        F1[scorer.py\ncatalog_score = text*0.6 + visual*0.4]
    end

    T2 --> F1
    V3 --> F1
    V5 --> F1
    F1 --> BQ[(product_scores)]

Loading

---

API endpoints

Method	Path	Description
GET	/health	Liveness check
GET	/catalog/summary	Global score averages
GET	/catalog/evolution	30-day score history
GET	/catalog/categories	Category ranking
GET	/products	Paginated product list with filters
GET	/products/{ean}	Full product detail
GET	/anomalies	Detected anomalies (7 days)
GET	/reports/latest	Latest Gemini report
POST	/search/visual	CLIP visual similarity search

---

Visual search

sequenceDiagram
    participant UI as Next.js
    participant API as FastAPI
    participant CLIP as CLIPModel
    participant BQ as BigQuery

    UI->>API: POST /search/visual {image_url, k}
    API->>API: fetch image (PIL RGB)
    API->>CLIP: manual preprocess (224x224 BICUBIC, CLIP norm)
    CLIP-->>API: embedding float[512]
    API->>BQ: VECTOR_SEARCH(clip_embeddings, embedding, top_k)
    BQ-->>API: [{ean, distance}]
    API->>BQ: JOIN product_scores for metadata
    API-->>UI: [{ean, product_name, similarity, thumbnail_url}]
    UI->>UI: render grid sorted by similarity

Loading

---

Scoring model

Dimension	Weight	Sub-components
Text score	60%	Name, brands, categories, nutriscore, quantity, packaging completeness
Visual score	40%	Resolution 40% + Sharpness 40% + Centration (Vision API) 20%

catalog_score = text_score * 0.6 + visual_score * 0.4

---

Tech stack

Layer	Technology
Frontend	Next.js 14, TypeScript, App Router, Vercel
API	FastAPI 0.111, Python 3.12, Cloud Run
Pipeline	Python, BigQuery ML, Cloud Build, Docker
AI / ML	CLIP (openai/clip-vit-base-patch32), Gemini 2.5 Flash, Vision API
Storage	BigQuery (scores, embeddings, reports), GCS (thumbnails)
Infra	GCP: Cloud Run, Cloud Build, Artifact Registry, Cloud Storage

---

Project structure

prisme/
├── api/                     # FastAPI service (Cloud Run)
│   ├── routers/             # catalog, products, anomalies, search, reports
│   ├── services/            # bigquery.py, clip.py, vertex.py
│   └── requirements.txt
├── pipeline/                # Daily audit pipeline (Cloud Build)
│   ├── encoder.py           # CLIP batch embedding to BQ
│   ├── scorer.py            # Fusion scoring
│   ├── report_generator.py  # Gemini AI daily report
│   └── ...
├── frontend/                # Next.js app (Vercel)
│   └── src/app/
│       ├── page.tsx         # Dashboard
│       ├── categories/      # Category ranking
│       ├── anomalies/       # Alert list
│       ├── products/        # Catalogue + product detail
│       ├── search/          # CLIP visual search
│       └── reports/         # Gemini AI reports
├── sql/                     # BigQuery table DDL + ARIMA views
├── infra/                   # BQ table creation, IAM, data loading
└── docker/                  # Dockerfile.api, Dockerfile.pipeline

---

Setup

Prerequisites

• GCP project with BigQuery, Cloud Run, Cloud Build, Vision API enabled
• Artifact Registry repository for Docker images
• Vercel account

Environment variables

Cloud Run (API)

GOOGLE_APPLICATION_CREDENTIALS or Workload Identity
GEMINI_API_KEY

Vercel (frontend)

NEXT_PUBLIC_API_URL=https://<cloud-run-url>

Deploy API

gcloud builds submit --config cloudbuild-api.yaml
gcloud run deploy prisme-api \
  --image europe-west1-docker.pkg.dev/<project>/prisme-docker/prisme-api:latest \
  --region europe-west1

Deploy frontend

cd frontend && vercel --prod

Run pipeline

docker build -f docker/Dockerfile.pipeline -t prisme-pipeline .
docker run --env-file .env prisme-pipeline