Cocktail AI: Evolutionary Recipe Generation with ML-Powered Flavor Prediction, deployed on AWS

Built an AI-assisted cocktail design system that predicts multi-label flavor profiles and uses evolutionary search to optimize recipes under real-world constraints. I trained taste classifiers on 4,500+ recipes, wired the models into a GA fitness loop, and deployed the pipeline on AWS for low-latency generation. The result is a production-style ML workflow that blends data science, backend orchestration, and DevOps to ship a user-facing product.

Average AUC: 0.90
Best AUC: 0.998
Average F1-Score: 0.67
Taste Classifiers: 9
DataSet: 4,500+

Core Innovation

What makes this project different:

Multi-label flavor modeling instead of single “primary taste.”
Ingredient quantities as features, not just presence/absence.
Genetic Algorithm for constrained search, not brute force.
Fitness driven by trained ML models, not hand-crafted rules.

This turns cocktail creation into a structured optimization problem similar to real-world engineering systems. It runs on a serverless AWS stack so it stays reliable, cost-aware, and responsive for users.

S3: stores datasets and trained models.
SageMaker: trains and serves the taste models.
SQS + Lambda: processes recipe jobs in the background.
DynamoDB: tracks job status for the UI.
CloudWatch: monitors health and performance.

Performance highlights: average final fitness of 95.2%, average fitness of 90.6%, and recipe generation completing in about 45 seconds end-to-end.

The system scales smoothly by batching model evaluations per generation, cutting inference calls by 99.9% across a full run while keeping quality intact.

AWS Architecture Diagram — Event-driven ingestion, model training, and low-latency API delivery.

Sample Generated Recipes

Each recipe is evolved from mandatory ingredients and target flavor profiles using ML-powered genetic optimization.

spicy 90%sweet 10%

Mandatory

Sugar Syrup (2:1), Vodka

Generated Composition

Chocolate Caramel Liqueurs35.00%

Ginger Liqueur31.51%

Vodka15.00%

Sugar Syrup (2:1)15.00%

Total normalized: 96.51%

sweet 80%herbal 20%

Mandatory

Lime Juice, Vodka

Generated Composition

Chocolate Caramel Liqueurs32.28%

Bénédictine D.O.M.23.99%

Vodka15.08%

Lime Juice15.00%

Advocaat Liqueur13.62%

Total normalized: 99.97%

floral 50%fruity 50%

Mandatory

Vodka, Spiced Rum

Generated Composition

Crème de Cassis35.00%

Crème de Violette32.61%

Spiced Rum15.00%

Vodka15.00%

Total normalized: 97.61%

citrus 100%

Mandatory

None

Generated Composition

Lemon Juice35.00%

Lime Juice35.00%

Limoncello Liqueur29.00%

Total normalized: 99.00%

creamy 60%sweet 40%

Mandatory

Gin

Generated Composition

Whipping Cream35.00%

Cream of Coconut35.00%

Gin24.53%

Total normalized: 94.53%

bittersweet 70%fruity 30%

Mandatory

Malt Whisky

Generated Composition

Fresh Fruit34.97%

Gentian Liqueur34.96%

Malt Whisky24.29%

Total normalized: 94.22%

Model Performance

All taste classifiers were trained using Logistic Regression with tuned thresholds:

Flavor	AUC	Precision	Recall	F1-Score
Creamy	0.998	0.917	1.000	0.957
Floral	0.971	0.458	0.647	0.537
Bittersweet	0.960	0.802	0.853	0.827
Spicy	0.943	0.571	0.600	0.585
Sweet	0.901	0.731	0.731	0.731
Herbal	0.881	0.531	0.682	0.597
Fruity	0.863	0.750	0.596	0.664
Citrus	0.849	0.660	0.865	0.749
Savoury	0.750	0.391	0.429	0.409

Even with simple linear models, the system achieves strong AUC and usable F1 scores — validating the feature design and framing of flavor prediction.

These models directly power the GA fitness evaluation.