/** * Delectable AI ML/Personalization Pipeline * * Interactive visualization of the actual Delectable AI models and * personalization pipeline powering the Your Grocer shopping assistant. * * Inline JSX + Babel pattern (no imports, no exports). */ const { useState, useEffect } = React; // ── Inline SVG Icon Components ────────────────────────────────────────────── const Database = () => ( ); const Brain = () => ( ); const Zap = () => ( ); const Target = () => ( ); const TrendingUp = () => ( ); const Users = () => ( ); const ShoppingCart = () => ( ); const GitBranch = () => ( ); const Layers = () => ( ); const Activity = () => ( ); const BarChart3 = () => ( ); const Network = () => ( ); const Sparkles = () => ( ); const ChevronRight = () => ( ); const Play = () => ( ); const Pause = () => ( ); const Filter = () => ( ); const Cpu = () => ( ); const Box = () => ( ); const ArrowRight = () => ( ); const Beaker = () => ( ); const Shield = () => ( ); const ShoppingBag = () => ( ); const Clock = () => ( ); const LineChart = () => ( ); // ── Main Component ────────────────────────────────────────────────────────── const MLPersonalizationPipeline = () => { const [activeStage, setActiveStage] = useState(0); const [isAnimating, setIsAnimating] = useState(false); const [selectedModel, setSelectedModel] = useState('dietary-propensity'); const [expandedCode, setExpandedCode] = useState(null); const stages = [ { id: 0, name: 'Data Sources', icon: Database }, { id: 1, name: 'Feature Engineering', icon: Cpu }, { id: 2, name: 'Model Training', icon: Brain }, { id: 3, name: 'Personalization', icon: Sparkles } ]; useEffect(() => { if (isAnimating) { const interval = setInterval(() => { setActiveStage((prev) => (prev + 1) % stages.length); }, 4000); return () => clearInterval(interval); } }, [isAnimating]); // ── Stage 0: Data Sources ────────────────────────────────────────────── const dataSources = [ { id: 'purchase-history', name: 'Purchase History (BigQuery)', icon: ShoppingCart, color: '#10b981', records: '48M', description: '2-year transaction data across all Your Grocer banners', features: ['Basket composition', 'Purchase frequency', 'Category affinity', 'Time patterns', 'Price sensitivity'], schema: { 'loyal_card_no': 'string', 'transaction_date': 'datetime', 'UPC': 'string', 'quantity': 'int', 'price': 'float', 'store_id': 'string' } }, { id: 'product-catalog', name: 'Product Catalog', icon: Box, color: '#8b5cf6', records: '70K+', description: 'Enriched product data from USDA, Open Food Facts, Tufts Food Compass', features: ['Name & brand', 'Category taxonomy', 'Nutrition facts', 'Ingredients list', 'Health labels (organic, GF, etc.)', 'UPC mapping'], schema: { 'UPC': 'string', 'product_name': 'string', 'brand': 'string', 'category': 'string', 'nutrition_json': 'json', 'ingredients': 'text', 'health_labels': 'array' } }, { id: 'behavioral', name: 'Behavioral Signals', icon: Activity, color: '#f59e0b', records: '1.2B', description: 'Clickstream, search queries, and cart events from web and app', features: ['Clickstream events', 'Search queries', 'Cart add/remove', 'Session patterns', 'Browse-to-buy signals'], schema: { 'session_id': 'string', 'event_type': 'enum', 'product_id': 'string', 'timestamp': 'datetime', 'query_text': 'string' } }, { id: 'loyalty', name: 'Loyalty & Demographics', icon: Users, color: '#ec4899', records: '135K', description: 'Shopper profiles with loyalty tiers, household info, and sensitivity scores', features: ['Loyalty tier', 'Household size', 'Price sensitivity', 'Promo sensitivity', 'Rewards balance'], schema: { 'loyal_card_no': 'string', 'loyalty_tier': 'enum', 'household_size': 'int', 'price_sensitivity': 'float', 'promo_sensitivity': 'float', 'rewards_balance': 'int' } } ]; // ── Stage 1: Feature Engineering ─────────────────────────────────────── const featureEngineering = [ { id: 'dietary-signals', name: 'Dietary Signal Extraction', description: 'Analyze 2-year purchase history to compute 11 dietary propensity scores per shopper. Each score represents the probability of a dietary preference.', icon: Shield, color: 'from-green-500 to-emerald-500', examples: ['organic_propensity_score', 'gluten_free_propensity', 'vegan_propensity', 'keto_propensity'], code: `# Dietary propensity extraction from purchase history def compute_dietary_propensities(purchases_2yr): """Compute 11 dietary probability scores from purchase patterns.""" total_spend = purchases_2yr.amount.sum() organic_ratio = purchases_2yr[purchases_2yr.is_organic].amount.sum() / total_spend gluten_free_ratio = purchases_2yr[purchases_2yr.is_gluten_free].amount.sum() / total_spend # 11 scores: vegan, vegetarian, gluten_free, dairy_free, nut_free, # keto, low_sodium, low_sugar, high_protein, high_fiber, organic scores = { "organic": organic_ratio, "gluten_free": gluten_free_ratio, "vegan": vegan_ratio, "keto": keto_ratio, # ... all 11 dietary dimensions } return { score: sigmoid(ratio * calibration_factor) for score, ratio in scores.items() }` }, { id: 'consumption-velocity', name: 'Consumption Velocity Estimation', description: 'Model per-product consumption rate to estimate virtual pantry stock levels. Predicts when a shopper will run out of each product.', icon: Clock, color: 'from-blue-500 to-cyan-500', examples: ['pantry_stock_olive_oil', 'consumption_velocity_milk', 'days_until_reorder_eggs', 'stock_level_0_to_1'], code: `# Virtual pantry stock estimation def estimate_stock_level(product, purchase_history): """Estimate remaining stock for a product-household pair.""" avg_purchase_interval = purchase_history.diff().mean() days_since_last = (now - last_purchase).days shelf_life = get_shelf_life(product.category) consumption_rate = product.typical_quantity / avg_purchase_interval remaining = max(0, last_quantity - consumption_rate * days_since_last) return min(1.0, remaining / product.typical_quantity) # Example output: # {"olive_oil": 0.73, "milk_2pct": 0.12, "eggs_dozen": 0.05} # \u2192 milk and eggs flagged for reorder` }, { id: 'basket-composition', name: 'Basket Composition Analysis', description: 'Extract co-purchase patterns and build a knowledge graph of product affinities from transaction baskets.', icon: Network, color: 'from-purple-500 to-pink-500', examples: ['co_purchase_affinity_score', 'basket_category_vector', 'mission_cluster_id', 'trip_type_label'], code: `# Co-purchase pattern extraction via association rules from mlxtend.frequent_patterns import apriori, association_rules # Build transaction matrix basket_matrix = ( df.groupby(["transaction_id", "product_name"])["quantity"] .sum().unstack().fillna(0) .applymap(lambda x: 1 if x > 0 else 0) ) # Mine frequent itemsets (min 0.5% support) frequent = apriori(basket_matrix, min_support=0.005, use_colnames=True) rules = association_rules(frequent, metric="lift", min_threshold=2.0) # Top co-purchase pairs feed the knowledge graph # e.g., {pasta \u2192 marinara_sauce, lift: 8.4}` }, { id: 'embeddings', name: 'Multi-Aspect Embedding Generation', description: 'Generate 3 types of 768-dimensional embeddings per product using Vertex AI text-embedding-005. Identity, nutrition, and description aspects.', icon: Layers, color: 'from-orange-500 to-amber-500', examples: ['embedding_identity_768d', 'embedding_nutrition_768d', 'embedding_description_768d', 'cosine_similarity_score'], code: `# Generate 3-aspect embeddings via Vertex AI from vertexai.language_models import TextEmbeddingModel model = TextEmbeddingModel.from_pretrained("text-embedding-005") def generate_product_embeddings(product): """Generate 3 x 768-dim embeddings per product.""" identity_text = f"{product.brand} {product.name} {product.category}" nutrition_text = f"Calories {product.calories} Fat {product.fat}g ..." description_text = product.description or product.name embeddings = model.get_embeddings([ identity_text, nutrition_text, description_text ]) return { "identity": embeddings[0].values, # 768-dim "nutrition": embeddings[1].values, # 768-dim "description": embeddings[2].values, # 768-dim } # Stored in BigQuery for VECTOR_SEARCH # 70K products x 3 aspects = 210K vectors` } ]; // ── Stage 2: Model Training ──────────────────────────────────────────── const mlModels = { 'dietary-propensity': { name: 'Dietary Propensity Model', type: 'Ensemble (Logistic Regression + calibrated probabilities)', description: 'Predicts 11 dietary preference probabilities per shopper from 2-year purchase history patterns.', input: '2yr purchase history features: category ratios, brand preferences, nutrition profile of basket', output: '11 probability scores (0\u20131) for each dietary preference', metrics: { 'AUC': '0.89', 'Calibration Error': '0.03', 'Coverage': '94%', 'Refresh Cadence': '24h' }, useCases: ['Dietary badge annotations', 'Search result filtering', 'Personalized warnings'], code: `# Dietary Propensity Model - Ensemble approach from sklearn.linear_model import LogisticRegression from sklearn.calibration import CalibratedClassifierCV DIETARY_DIMENSIONS = [ "vegan", "vegetarian", "gluten_free", "dairy_free", "nut_free", "keto", "low_sodium", "low_sugar", "high_protein", "high_fiber", "organic" ] def train_dietary_model(features_df, labels_df): """Train calibrated logistic regression for each dietary dim.""" models = {} for dim in DIETARY_DIMENSIONS: base = LogisticRegression( C=1.0, max_iter=1000, class_weight="balanced" ) calibrated = CalibratedClassifierCV(base, cv=5, method="isotonic") calibrated.fit(features_df, labels_df[dim]) models[dim] = calibrated return models # Output per shopper: # {"organic": 0.82, "gluten_free": 0.78, "keto": 0.15, ...}` }, 'virtual-pantry': { name: 'Virtual Pantry Estimator', type: 'Time-series regression per product-household', description: 'Estimates per-product stock levels for each household, predicting when items will need replenishment.', input: 'Purchase timestamps, quantities, product shelf life, household size', output: 'Estimated stock level (0\u20131), days until reorder, consumption velocity', metrics: { 'MAE': '0.12', 'Stock-out Accuracy': '84%', 'Products Tracked': '2,400', 'Refresh Cadence': '24h' }, useCases: ['Shopping list auto-exclusion', 'Reorder reminders', 'Low-stock flagging'], code: `# Virtual Pantry Estimator - Time-series regression import numpy as np from scipy.optimize import curve_fit def estimate_consumption_rate(purchase_history): """Fit consumption curve per product-household pair.""" intervals = purchase_history["date"].diff().dt.days.dropna() quantities = purchase_history["quantity"].values # Exponential decay model for stock depletion avg_interval = intervals.mean() avg_quantity = quantities.mean() consumption_rate = avg_quantity / avg_interval # units per day return { "consumption_rate": consumption_rate, "avg_interval_days": avg_interval, "avg_quantity": avg_quantity, } def get_stock_level(product_id, household_id): """Current estimated stock as fraction of typical purchase.""" rate = consumption_rates[(product_id, household_id)] days_elapsed = (now - last_purchase_date).days remaining = max(0, last_qty - rate * days_elapsed) return min(1.0, remaining / typical_purchase_qty) # stock > 0.5 \u2192 exclude from list # stock 0.1-0.5 \u2192 flag "running low" # stock < 0.1 \u2192 add to reorder suggestions` }, 'mission-cluster': { name: 'Mission Cluster Model', type: 'K-means clustering (k=6)', description: 'Segments shopping trips into 6 mission types based on basket composition, trip frequency, and behavioral patterns.', input: 'Basket composition vectors, trip frequency, spend patterns, time-of-day', output: '6 clusters: Weekly Stock-Up, Quick Meal, Health Focus, Budget Batch, Special Occasion, Exploration', metrics: { 'Silhouette Score': '0.72', 'Clusters': '6', 'Avg Cluster Size': '22.5K', 'Stability': '0.89' }, useCases: ['Mission-targeted promotions', 'UI personalization', 'Inventory planning'], code: `# Mission Cluster Model - K-means on trip features from sklearn.cluster import KMeans from sklearn.preprocessing import StandardScaler MISSION_LABELS = { 0: "Weekly Stock-Up", 1: "Quick Meal", 2: "Health Focus", 3: "Budget Batch", 4: "Special Occasion", 5: "Exploration", } def train_mission_clusters(trip_features): """Cluster shopping trips into 6 mission types.""" scaler = StandardScaler() X = scaler.fit_transform(trip_features[[ "basket_size", "unique_categories", "avg_item_price", "pct_produce", "pct_protein", "pct_snacks", "time_of_day", "day_of_week", "trip_frequency_30d", ]]) kmeans = KMeans(n_clusters=6, random_state=42, n_init=10) kmeans.fit(X) trip_features["mission"] = [ MISSION_LABELS[c] for c in kmeans.labels_ ] return kmeans, scaler # "Health Focus" cluster \u2192 organic deals # "Budget Batch" cluster \u2192 bulk discounts` }, 'embedding-generator': { name: 'Multi-Aspect Embedding Generator', type: 'Vertex AI text-embedding-005', description: 'Generates 3 semantic vector representations per product for identity, nutrition, and description search.', input: 'Product text (3 aspects: identity, nutrition, description)', output: '3 \u00d7 768-dimensional vectors per product, stored in BigQuery', metrics: { 'Dimensions': '768', 'Total Vectors': '210K', 'Index Type': 'IVF', 'Search Latency': '<50ms' }, useCases: ['Semantic product search', 'Similar product discovery', 'Natural language queries'], code: `# Multi-Aspect Embedding Generation + BigQuery VECTOR_SEARCH # 1. Generate embeddings (batch job, nightly) from vertexai.language_models import TextEmbeddingModel model = TextEmbeddingModel.from_pretrained("text-embedding-005") for product in product_catalog: embeddings = model.get_embeddings([ f"{product.brand} {product.name} {product.category}", format_nutrition_text(product.nutrition), product.description or product.name, ]) store_embeddings(product.upc, embeddings) # 2. Query via BigQuery VECTOR_SEARCH at inference time VECTOR_SEARCH_SQL = """ SELECT base.product_name, base.brand, base.UPC, distance FROM VECTOR_SEARCH( TABLE \`project.dataset.product_embeddings\`, 'embedding_identity', (SELECT embedding FROM query_embedding), top_k => 20, distance_type => 'COSINE' ) """ # Returns top-20 semantically similar products in <50ms` }, 'food-intelligence': { name: 'Food Intelligence Scorer', type: 'Rule-based + algorithmic (Food Compass, Nutri-Score, NOVA)', description: 'Computes nutritional quality scores for products using established food science algorithms.', input: 'Nutrition facts, ingredients list, processing indicators', output: 'Food Compass (1\u2013100), Nutri-Score (A\u2013E), NOVA (1\u20134)', metrics: { 'Food Compass': '1\u2013100 scale', 'Nutri-Score': 'A\u2013E', 'NOVA': '1\u20134', 'Coverage': '42%' }, useCases: ['Healthier swap suggestions', 'Ultra-processed warnings', 'Nutrition-aware ranking'], code: `# Food Intelligence Scoring Pipeline def compute_food_compass(nutrition, ingredients): """Tufts Food Compass score (1-100, higher = healthier).""" # 54 attributes across 9 domains score = ( nutrient_ratio_domain(nutrition) + vitamin_mineral_domain(nutrition) + food_ingredient_domain(ingredients) + additive_domain(ingredients) + processing_domain(ingredients) + lipid_domain(nutrition) + fiber_protein_domain(nutrition) + phytochemical_domain(ingredients) + specific_lipid_domain(nutrition) ) return clamp(score, 1, 100) def compute_nutri_score(nutrition): """EU Nutri-Score: A (best) to E (worst).""" negative = (nutrition.energy_pts + nutrition.sugar_pts + nutrition.saturated_fat_pts + nutrition.sodium_pts) positive = (nutrition.fiber_pts + nutrition.protein_pts + nutrition.fruit_veg_pts) score = negative - positive return map_to_letter(score) # A-E def compute_nova(ingredients): """NOVA classification: 1=unprocessed to 4=ultra-processed.""" if has_ultra_processed_markers(ingredients): return 4 # ... classification logic return nova_class` }, 'co-purchase-graph': { name: 'Co-Purchase Knowledge Graph', type: 'Association rule mining + graph embedding', description: 'Builds a product affinity graph from transaction baskets for basket completion and recipe mapping.', input: 'Transaction baskets (product co-occurrence)', output: 'Product affinity scores, "frequently bought together" edges', metrics: { 'Edges': '1.2M', 'Avg Lift': '3.4', 'Min Support': '0.5%', 'Update Freq': 'Weekly' }, useCases: ['Basket completion suggestions', 'Recipe ingredient mapping', '"Frequently bought together"'], code: `# Co-Purchase Knowledge Graph Construction from mlxtend.frequent_patterns import apriori, association_rules import networkx as nx # 1. Mine association rules from transaction baskets frequent_items = apriori( basket_matrix, min_support=0.005, use_colnames=True ) rules = association_rules( frequent_items, metric="lift", min_threshold=2.0 ) # 2. Build directed graph of product affinities G = nx.DiGraph() for _, rule in rules.iterrows(): for ant in rule["antecedents"]: for con in rule["consequents"]: G.add_edge(ant, con, weight=rule["lift"], confidence=rule["confidence"]) # 3. Query: "What goes with pasta?" neighbors = sorted( G[\"pasta\"].items(), key=lambda x: x[1]["weight"], reverse=True )[:5] # \u2192 [("marinara_sauce", 8.4), ("parmesan", 6.2), ...] # 4. Recipe ingredient mapping def suggest_basket_completion(current_basket, recipe_graph): missing = recipe_graph.required - set(current_basket) return [(item, G[nearest][item]["weight"]) for item in missing if item in G]` } }; // ── Stage 3: Personalization Outputs ─────────────────────────────────── const personalizationOutputs = [ { type: 'Dietary-Aware Search Results', icon: Shield, description: 'Agent annotates products with dietary badges based on propensity scores > 0.5. A shopper with gluten_free=0.78 sees gluten warnings prominently.', model: 'Dietary Propensity Model', example: 'Shopper (gluten_free: 0.78) searches "pasta" \u2192 Results show GF badge on rice pasta, warning on wheat pasta. Organic items boosted for organic_propensity > 0.6.', color: 'from-green-500 to-emerald-600' }, { type: 'Virtual Pantry Deductions', icon: ShoppingBag, description: 'Shopping lists automatically exclude items with stock > 0.5 and flag items at 0.1\u20130.5. The optimize_shopping_list tool handles this deterministically.', model: 'Virtual Pantry Estimator', example: 'Recipe calls for olive oil (stock: 0.73) \u2192 excluded. Milk (stock: 0.12) \u2192 flagged "running low". Eggs (stock: 0.05) \u2192 added to list.', color: 'from-blue-500 to-cyan-600' }, { type: 'Mission-Targeted Promotions', icon: Target, description: 'Promos matched to the shopper\'s detected mission cluster. "Health Focus" cluster gets organic deals; "Budget Batch" sees bulk pricing.', model: 'Mission Cluster Model', example: 'Shopper in "Health Focus" cluster \u2192 Agent surfaces organic produce deals, plant-based protein promos, and wellness bundle discounts.', color: 'from-purple-500 to-pink-600' }, { type: 'Personalized Product Ranking', icon: TrendingUp, description: 'Search results re-ranked by combining dietary fit, purchase affinity, food intelligence scores, and semantic similarity.', model: 'Multi-Aspect Embeddings + Food Intelligence', example: 'Query "healthy snack" \u2192 Re-ranked: (1) Almonds [Food Compass: 88, organic_fit: 0.9], (2) Greek Yogurt [FC: 76], (3) Granola Bar [NOVA: 4, flagged].', color: 'from-orange-500 to-amber-600' }, { type: 'Shopping List Optimization', icon: ShoppingCart, description: 'Cross-recipe ingredient aggregation, unit conversion, and retail package mapping. All deterministic via the optimize_shopping_list tool.', model: 'Co-Purchase Graph + Virtual Pantry', example: '3 recipes need "garlic" \u2192 aggregated to 1 bulb. "400g chicken" \u2192 mapped to "1 lb pack". Missing "basil" detected from co-purchase graph.', color: 'from-teal-500 to-green-600' }, { type: 'Healthier Alternative Suggestions', icon: Beaker, description: 'Food Compass scores drive "healthier swap" recommendations. NOVA flags ultra-processed items. Agent suggests alternatives proactively.', model: 'Food Intelligence Scorer + Embeddings', example: 'Cart has instant ramen (NOVA: 4, FC: 22) \u2192 Agent suggests fresh ramen noodles (NOVA: 1, FC: 67) with similar embeddings, +45 Food Compass points.', color: 'from-rose-500 to-red-600' } ]; // ── Render ───────────────────────────────────────────────────────────── return (
{/* ── Header ──────────────────────────────────────────────────────── */}

Delectable AI ML/Personalization Pipeline

Your Grocer Shopping Intelligence

{/* ── Stage Navigation ────────────────────────────────────────────── */}
{stages.map((stage, idx) => { const Icon = stage.icon; const isActive = activeStage === stage.id; const isPast = activeStage > stage.id; return ( {idx < stages.length - 1 && (
)} ); })}
{/* ── Stage 0: Data Sources ───────────────────────────────────── */} {activeStage === 0 && (

Data Sources

Actual data feeding Delectable AI models

{dataSources.map((source) => { const Icon = source.icon; return (

{source.name}

{source.records} records

{source.description}

Features

{source.features.map((feature) => ( {feature} ))}

Schema

{Object.entries(source.schema).map(([field, type]) => (
{field} {type}
))}
); })}
{/* Data Pipeline Architecture */}

Data Pipeline Architecture

{[ { icon: Database, label: 'BigQuery', sub: 'Raw data' }, { icon: Filter, label: 'dbt Transforms', sub: 'Clean & join' }, { icon: Cpu, label: 'Feature Extraction', sub: 'Signals & scores' }, { icon: Brain, label: 'Model Training', sub: 'Vertex AI' }, { icon: Zap, label: 'Serving', sub: 'Real-time API' }, ].map((step, i) => { const StepIcon = step.icon; return (

{step.label}

{step.sub}

 ); })}
BigQuery \u2192 dbt \u2192 Feature Store \u2192 Vertex AI \u2192 Agent Tools
)} {/* ── Stage 1: Feature Engineering ────────────────────────────── */} {activeStage === 1 && (

Feature Engineering

Delectable AI-specific signal extraction from raw data

{featureEngineering.map((feature) => { const Icon = feature.icon; const isExpanded = expandedCode === feature.id; return (

{feature.name}

{feature.description}

Example Features

{feature.examples.map((ex) => ( {ex} ))}
{isExpanded && ( 
{feature.code}
)}
); })}
{/* Feature Summary */}

Feature Pipeline Summary

11

Dietary Scores

2,400

Pantry Products

210K

Embedding Vectors

6

Mission Clusters

)} {/* ── Stage 2: Model Training ─────────────────────────────────── */} {activeStage === 2 && (

Delectable AI Models

Actual models powering the personalization pipeline

{/* Model Selector */}
{Object.entries(mlModels).map(([key, model]) => ( ))}
{/* Selected Model Details */} {selectedModel && mlModels[selectedModel] && (() => { const model = mlModels[selectedModel]; return (

{model.name}

{model.type}

{model.description}

{/* Left: Input/Output & Metrics */}

Input

{model.input}

Output

{model.output}

Performance Metrics

{Object.entries(model.metrics).map(([metric, value]) => (

{value}

{metric}

))}

Use Cases

{model.useCases.map((uc) => (
{uc}
))}
{/* Right: Code */}

Implementation

{model.code}
); })()} {/* Training Pipeline */}

Training & Serving Pipeline

{[ { icon: Database, label: 'BigQuery', sub: 'Feature tables' }, { icon: GitBranch, label: 'Train/Val Split', sub: '80/20 temporal' }, { icon: Brain, label: 'Vertex AI Training', sub: 'Managed training' }, { icon: BarChart3, label: 'Evaluation', sub: 'AUC, MAE, silhouette' }, { icon: Zap, label: 'Vertex AI Serving', sub: 'Online prediction' } ].map((step, i) => { const StepIcon = step.icon; return (

{step.label}

{step.sub}

); })}
)} {/* ── Stage 3: Personalization Outputs ────────────────────────── */} {activeStage === 3 && (

Personalization Outputs

How Delectable AI models feed the shopping assistant agent

{personalizationOutputs.map((output) => { const Icon = output.icon; return (

{output.type}

{output.description}

{output.model}

Example

{output.example}

); })}
{/* Production Metrics */}

Production Metrics

24h

Profile Refresh

Stale-While-Revalidate

210K

Embedding Vectors

70K x 3 aspects

~4,200

Tool Calls/Day

Agent invocations

94%

Personalization Rate

Responses w/ signal

2.8s

Avg Response Time

Including tool calls

{/* Context Engineering Highlight */}

How Models Feed the Agent

Delectable AI uses a tool-calling architecture where the LLM agent invokes deterministic tools enriched by ML model outputs. Models pre-compute shopper signals; the agent consumes them at inference time.

Shopper Profile (Pre-computed)

  • - 11 dietary propensity scores
  • - Mission cluster assignment
  • - Virtual pantry stock levels
  • - Purchase frequency patterns

Product Intelligence (Indexed)

  • - 3 x 768d embedding vectors
  • - Food Compass / Nutri-Score / NOVA
  • - Co-purchase affinity edges
  • - Health label classifications

Agent Tools (Deterministic)

  • - search_products (vector + keyword)
  • - optimize_shopping_list
  • - get_shopper_profile
  • - lookup_promotions

End-to-end flow:{' '} Shopper profile loaded (cached, 24h SWR) → Agent receives query + profile context → Tool calls with profile-enriched parameters → Results ranked by dietary fit + food intelligence → Personalized response in ~2.8s

)}
{/* ── Footer ──────────────────────────────────────────────────────── */}

Delectable AI ML/Personalization Pipeline • Your Grocer

Powered by Vertex AI, BigQuery ML, dbt, and the Delectable AI Agent Framework

{/* Inline animation styles */}
); }; // ── Render ──────────────────────────────────────────────────────────────── const root = ReactDOM.createRoot(document.getElementById('root')); root.render();