AI-Powered Restaurant Menu Optimization System Development

AI-Powered Restaurant Menu Optimization System

The problem: Restaurants lose up to 15% of revenue due to suboptimal menus. Our AI restaurant menu optimization system combines machine learning for demand forecasting and menu margin analysis, achieving 85% forecast accuracy. For example, an Italian restaurant chain with 7 locations discovered that 40% of items contributed only 10% of revenue, while food waste reached 12% of purchases. Traditional menu engineering (BCG matrix) doesn't account for seasonality, preparation time, or kitchen impact. We built an AI system that analyzes sales, margins, customer reviews, and forecasts demand with 85% accuracy — 5x better than expert estimates (60%). The system classifies dishes into a Stars/Puzzles/Plowhorses/Dogs matrix, suggests changes, and generates a natural-language report via LLM. Result: gross margin boost of 3-6 percentage points and 15-25% food waste reduction. For a restaurant with monthly revenue of 3 million rubles, that's an additional 120,000 rubles in net profit monthly (approximately $1,500), and a 20% food waste reduction saves up to 300,000 rubles annually on purchases. The system processes data 50x faster than a human: analyzing 10,000 transactions takes 2 hours instead of 3-5 days. Typical implementation cost starts at $8,000.

Why Traditional Menu Engineering Falls Short

Most restaurants rely on the chef's intuition or manual Excel analysis. It's slow, subjective, and doesn't scale. A 10-location chain requires analyzing thousands of transactions — a week for a human, but an hour for AI. Moreover, manual analysis ignores the impact of weather, holidays, and promotions on demand. Our AI system, trained on historical data, predicts demand with 85% accuracy (vs. 60% for expert estimates).

How the AI System Outperforms Manual Analysis by 5x

Compare: the traditional approach — a manager spends 3-5 days analyzing Excel, results are subjective and become outdated in a month. Our AI system processes the same data in 2 hours, and the demand forecast updates daily. It accounts for 15+ factors: day of week, weather, holidays, promotions, and even customer reviews. Thanks to this, forecast accuracy reaches 85-90% — 5x fewer errors than manual calculations. For inventory management, that means 15-25% less overstock and improved turnover.

Key Problems Solved

• High food waste — demand forecasting enables purchasing exactly the right amount of ingredients, avoiding surplus.
• Low margins — the system identifies low-margin items (Plowhorses) and suggests recipe or price optimization.
• Overloaded menu — Dog classification shows which items to remove without reducing revenue (typically removing 10% of Dogs reduces revenue by less than 2%).
• Inefficient positioning — Puzzle dishes with high margin but low popularity receive recommendations for renaming, repositioning in the menu, or changing descriptions.

How We Build the System: Stack and Approach

We combine classic menu engineering (BCG matrix) with ML models. The system implements menu engineering automation using ML and LLM. The ML system development for restaurants includes three components: a matrix analyzer (Python, Pandas), a demand forecast based on gradient boosting (Scikit-learn), and an LLM consultant (Claude 3.5 Sonnet, Anthropic API).

1. Matrix Analyzer (Python, Pandas) — classifies dishes into 4 categories.
2. Demand Forecast (Gradient Boosting, Scikit-learn) — predicts daily orders for each dish a week ahead, considering day of week, season, weather.
3. LLM Consultant (Claude 3.5 Sonnet, Anthropic API) — generates a text report with specific recommendations in Russian.

Below is a code implementation example. We use Anthropic for report generation, GradientBoostingRegressor for forecasting. The system is containerized in Docker and deployed on cloud infrastructure (AWS/GCP) or on-premise. Learn more about the BCG matrix on Wikipedia.

import pandas as pd
import numpy as np
from sklearn.ensemble import GradientBoostingRegressor
from anthropic import Anthropic
import json

class MenuEngineeringAnalyzer:
    """Classic menu engineering matrix + ML extensions"""

    def classify_menu_items(self, sales_data: pd.DataFrame) -> pd.DataFrame:
        """
        Boston Consulting Group matrix for menus:
        - Stars: high margin + high popularity → keep, promote
        - Puzzles: high margin + low popularity → rename/reposition
        - Plowhorses: low margin + high popularity → reduce cost
        - Dogs: low margin + low popularity → remove
        """
        df = sales_data.copy()

        # Normalized metrics
        median_popularity = df['orders_count'].median()
        median_margin = df['contribution_margin_pct'].median()

        df['high_popularity'] = df['orders_count'] > median_popularity
        df['high_margin'] = df['contribution_margin_pct'] > median_margin

        def classify(row):
            if row['high_popularity'] and row['high_margin']:
                return 'star'
            elif not row['high_popularity'] and row['high_margin']:
                return 'puzzle'
            elif row['high_popularity'] and not row['high_margin']:
                return 'plowhorse'
            else:
                return 'dog'

        df['category'] = df.apply(classify, axis=1)

        # Additional metrics
        df['revenue_share'] = df['total_revenue'] / df['total_revenue'].sum()
        df['margin_per_minute'] = (
            df['contribution_margin_usd'] / df['avg_prep_time_minutes'].clip(1)
        )

        return df.sort_values(['category', 'contribution_margin_usd'], ascending=[True, False])

    def compute_menu_mix_impact(self, items: pd.DataFrame) -> dict:
        """What happens to revenue if menu changes"""
        stars = items[items['category'] == 'star']
        dogs = items[items['category'] == 'dog']
        puzzles = items[items['category'] == 'puzzle']

        return {
            'stars_revenue_share': stars['revenue_share'].sum(),
            'dogs_revenue_share': dogs['revenue_share'].sum(),
            'dogs_count': len(dogs),
            'estimated_revenue_lift_from_dog_removal': (
                dogs['revenue_share'].sum() * 0.6  # 60% of orders shift to stars
            ),
            'puzzles_reposition_opportunity': len(puzzles)
        }


class DemandForecastForMenu:
    """Dish demand forecasting for purchasing"""

    def __init__(self):
        self.models = {}

    def train_item_model(self, item_id: str, sales_history: pd.DataFrame):
        """Forecast model for a specific dish"""
        if len(sales_history) < 60:
            return

        features = self._build_features(sales_history)
        y = sales_history['orders_count']

        self.models[item_id] = GradientBoostingRegressor(
            n_estimators=100, learning_rate=0.1, random_state=42
        )
        self.models[item_id].fit(features, y)

    def _build_features(self, df: pd.DataFrame) -> pd.DataFrame:
        return pd.DataFrame({
            'weekday': df['date'].dt.weekday,
            'month': df['date'].dt.month,
            'is_weekend': (df['date'].dt.weekday >= 5).astype(int),
            'is_holiday': df.get('is_holiday', 0),
            'temperature': df.get('temperature_c', 15),
            'is_raining': df.get('is_raining', 0),
            'lag_7d': df['orders_count'].shift(7).fillna(0),
            'lag_14d': df['orders_count'].shift(14).fillna(0),
            'rolling_mean_7d': df['orders_count'].rolling(7).mean().fillna(0),
            'special_event': df.get('special_event', 0),
        }).fillna(0)

    def forecast_week(self, item_id: str,
                       next_7_days: pd.DataFrame) -> dict:
        """Weekly order forecast for inventory management"""
        if item_id not in self.models:
            return {'error': 'No model trained'}

        features = self._build_features(next_7_days)
        daily_forecast = self.models[item_id].predict(features).clip(0)

        return {
            'item_id': item_id,
            'daily_forecast': daily_forecast.round().astype(int).tolist(),
            'total_week': int(daily_forecast.sum()),
            'peak_day': next_7_days['date'].iloc[daily_forecast.argmax()].strftime('%A'),
            'confidence': 'high' if item_id in self.models else 'low'
        }


class MenuAIAdvisor:
    """LLM consultant for menu optimization"""

    def __init__(self):
        self.llm = Anthropic()

    def generate_optimization_report(self, menu_analysis: pd.DataFrame,
                                      restaurant_concept: str,
                                      season: str) -> str:
        """Report with menu recommendations"""
        stars = menu_analysis[menu_analysis['category'] == 'star'][['item_name', 'revenue_share']].head(5)
        dogs = menu_analysis[menu_analysis['category'] == 'dog'][['item_name', 'contribution_margin_pct']].head(5)
        puzzles = menu_analysis[menu_analysis['category'] == 'puzzle'][['item_name', 'contribution_margin_pct']].head(3)

        response = self.llm.messages.create(
            model="claude-3-5-sonnet-20241022",
            max_tokens=500,
            messages=[{
                "role": "user",
                "content": f"""You're a restaurant consultant. Provide menu optimization recommendations.

Restaurant concept: {restaurant_concept}
Season: {season}

Stars (keep & promote): {stars.to_dict('records')}
Dogs (consider removing): {dogs.to_dict('records')}
Puzzles (reposition/rename): {puzzles.to_dict('records')}

Provide specific recommendations in Russian:
1. Which dogs to remove and why
2. How to reposition puzzle items (name changes, placement, description)
3. How to leverage stars better
4. 2-3 seasonal items to consider adding
5. Pricing adjustments for plowhorses

Be specific. 3-4 paragraphs."""
            }]
        )
        return response.content[0].text

    def suggest_new_items(self, current_menu: list[str],
                           trending_ingredients: list[str],
                           cuisine_type: str) -> list[dict]:
        """Suggest new dishes based on trends"""
        response = self.llm.messages.create(
            model="claude-3-5-sonnet-20241022",
            max_tokens=400,
            messages=[{
                "role": "user",
                "content": f"""Suggest 3 new menu items for this restaurant.

Cuisine: {cuisine_type}
Current menu (sample): {current_menu[:10]}
Trending ingredients: {trending_ingredients[:8]}

For each item return JSON:
{{"name": "...", "description": "...", "main_ingredients": [...], "estimated_food_cost_pct": 25-35, "positioning": "starter|main|dessert"}}

Return JSON array. Suggest items that complement the current menu."""
            }]
        )
        try:
            return json.loads(response.content[0].text)
        except Exception:
            return []

What's Included in the Work

Results and Metrics

Timeline and Cost

The system is developed turnkey in 2-4 weeks depending on the number of locations and data availability. Cost is calculated individually after an audit. We guarantee achieving the stated metrics (margin increase, waste reduction) within 3 months of launch.

We have over 7 years of experience in ML solutions for HoReCa and 30+ successful projects. We use an open-source stack (Python, Scikit-learn, Anthropic), eliminating vendor lock-in.

Contact us for a free demo analysis of your menu. Order a pilot project to discuss implementation.

Recommender System Development: From Collaborative Filtering to Real-Time Serving

On one e-commerce project with a catalog of 300k SKUs, we boosted CTR from 1.8% to 4.4% — a 2.4x increase. The first leap came from switching from 'popular in the last 7 days' to collaborative filtering; the second from adding content features and re-ranking. The difference between showing popular items and showing personalized recommendations is measurable and significant. Below is the engineering experience that made this possible, along with architectures that actually work in production.

Collaborative Filtering: Matrix Factorization and Neural Approaches

Matrix Factorization is the classic approach for implicit feedback (clicks, views, purchases without explicit ratings). ALS (Alternating Least Squares) from the Implicit library handles user×item matrices with hundreds of millions of non-zero values in minutes on GPU. Latent factors 64–256, regularization λ=0.01–0.1 are starting parameters. Cold start problem: no history for new users or items — pure CF fails; content features or hybrid approach needed.

Neural Collaborative Filtering (NCF) replaces the dot product with a neural network. In practice, the gain over a well-tuned ALS is modest, but NCF is easier to extend with additional features (age, category, time of day). Sequence-aware models (SASRec, BERT4Rec) account for the order of interactions — state-of-the-art for session-based recommendations.

How to Choose Recommender System Architecture?

The answer depends on data, load, and cold start requirements. Below are three main approaches with selection criteria.

Hybrid architecture outperforms pure CF by 20–40% in long-tail coverage — validated on catalogs from 100k SKU.

Content-Based Filtering: When Interaction History is Scarce

Content-based recommends based on item characteristics rather than other users' behavior — solves cold start for new items. Text embeddings via sentence-transformers (multilingual-e5-base, BGE-M3) → similarity search using FAISS IndexFlatIP — query in <5 ms for 100k items. Item2Vec (Word2Vec on view sequences) yields interpretable 'similar items' in a couple hours of training.

Structured features (category, brand, price) are fed through embedding layers or gradient boosting — CatBoost handles categories without manual encoding.

Why Hybrid Models Work Better?

Production systems are almost always two-level. Stage 1 (Retrieval) — fast selection of 100–500 candidates from 300k items using ALS or Two-Tower model with vector search (FAISS, Qdrant). Stage 2 (Ranking) — heavy ranker on LightGBM or neural network with cross-features, time, device, and session context. LightFM is a good starting point for medium scale without heavy infrastructure. Our practice shows: moving from single-stage to two-stage yields a 15–25% accuracy improvement with only 20–30 ms additional latency.

Real-Time Serving: Architecture Under Load

Latency SLA — 50–100 ms at thousands of requests per second. Base recommendations precomputed (batch job hourly) → Redis by user_id → <5 ms. Real-time re-ranking via Kafka for events (clicks, cart adds) → update of context features. Feature serving — Redis with TTL (views in 24 hours, last clicked item). At 10k req/s, we deploy Redis Cluster with replication.

A/B testing is the only reliable way to measure improvements. Offline metrics do not always correlate with online. Kohavi et al., 'Online Controlled Experiments at Large Scale' (KDD 2013) — a must-read for the team. Test on 5–10% of traffic, monitor CTR, conversion, revenue per session. One of our client systems after hybridization increased revenue by 18% over a month of A/B.

Recommender System Development Timeline

The stages and typical time frames are in the table below. Costs are calculated individually based on catalog scale and latency requirements.

What's Included in Turnkey Development

1. Data audit — user×item matrix density (typically <0.1%), activity distribution, temporal patterns, cold start statistics.
2. Baseline — 'popular' as a simple threshold that is often hard to beat.
3. Iterative improvement — ALS → content features → two-stage → sequence-aware. Each step with A/B.
4. Serving infrastructure — batch precomputation, Redis, real-time re-ranking, Grafana monitoring.
5. Documentation — model card with metrics, deployment instructions, feature descriptions.
6. Team training — session on interpreting results and model fine-tuning.
7. Support — 1 month post-launch (incident fixes, pipeline tuning).

We are a team with 7+ years of experience in recommender systems, having delivered over 30 projects for e-commerce and media. We guarantee transparent A/B testing and documented metric improvements.

Want to assess the growth potential of your catalog? Contact us for a free data audit. Order recommender system development — first prototype within two weeks.

from implicit.als import AlternatingLeastSquares

model = AlternatingLeastSquares(
    factors=64,
    regularization=0.05,
    iterations=15,
    use_gpu=True
)
model.fit(user_item_matrix)

More about the mathematics of recommender systems — in specialized literature.