Goulash extends an ML recommendation engine on AWS

Goulash is a Czech meal planning and grocery shopping app designed to make home cooking effortless. Powered by AI, it creates personalized weekly menus, chooses recipes based on tastes and preferences, calculates the right ingredients, auto-generates shopping lists, and prepares a grocery order through partners — leading Czech online grocery retailers — allowing users to go from meal idea to doorstep delivery in just a few taps. With a focus on reducing food waste and simplifying ingredient discovery, Goulash targets households looking to cut down on the daily effort involved in meal planning and grocery shopping. No wonder Goulash was named one of the top 15 innovations of 2025/2026 in the Ipsos Inspirational Innovations 2026 report.

Opportunity From data to dinner: analytics as the engine of personalization

The app has grown quickly — driven by influencer partnerships, integrations with leading Czech and German grocery chains, collaborations with top chefs, and strong organic traction — Goulash has built a loyal community and sustained growth of 70% month over month. But rapid growth creates a different challenge: retaining users by making every recommendation more relevant than the last. The team understood a simple principle: personalization is not a feature — it is a competitive advantage. The more accurately the platform understands how each user cooks, the more valuable the app becomes.

Goulash already operated a content-based recommendation system that matched dishes to explicit user preferences and previously selected meals. That system performs well for exactly what it was designed to do: it learns each user's stated tastes, ingredients and dietary preferences from the very first session, finds dishes similar to recipes a user has already liked or selected, excels at cold-start cases — new users with limited behavioral history and new recipes without interaction data yet — and composes varied meal plans with every generation. One capability, however, was deliberately outside its design scope: learning broader behavioral trends across the user base, and tracking how individual tastes shift over time. As the volume of behavioral data increased, Goulash wanted to extend its existing personalization capabilities with an additional behavioral signal and determine whether this could produce a measurable improvement in recommendation quality.

The company needed to determine whether a complementary approach — one that additionally learns from what users actually do in the app — could meaningfully improve recipe relevance, drive higher user satisfaction, and translate directly into increased frequency of use and stronger long-term retention. At the same time, Goulash was rightly cautious: rather than building first and validating later, Goulash's team chose the opposite approach. They turned to Chaos Gears, an AWS Premier Tier Partner with deep cloud and AI expertise and a proven track record of building data analytics platforms that turn user interactions into personalized, high-impact product experiences.

From the outset, the client's key challenges were clear:

• rapidly validating the potential of the new approach without a large budget commitment,
• laying the groundwork for future development of personalization features,
• ensuring the ability to measure recommendation quality and its actual impact on user behavior,
• maintaining full control over data and ensuring compliance with European data processing regulations.

These were business questions first. Technology needed to provide evidence — not assumptions. With extensive experience in building data analytics platforms that turn user interactions into actionable intelligence, Chaos Gears was well placed to help Goulash answer the right questions before making the wrong investments.

At Chaos Gears, we understand that ensuring data security is a critical task — just as building a competitive advantage through proper data analysis and quality is essential to meeting rising user demands while avoiding wasted budgets on ineffective experiments.

Solution An MLOps pipeline built to answer real questions

Goulash's recommendation feature was already in place and effective, taking explicit preferences into account and performing well for new users and new recipes. What Chaos Gears delivered was an additional engine within that feature: built on the same data, but generating recommendations through a fundamentally different, behavioral approach that learns how user tastes evolve — designed to work alongside the existing content-based logic, with each engine solving the problem it was designed to solve.

The goal was deliberately constrained — rather than architecting a production-grade recommendation system from the outset, we set out to build a complete MLOps pipeline that could take Goulash's raw user interaction data, process it, train a model, and return results that any engineer on the Goulash team could interpret and act on — no machine learning background required.

The starting point was the data itself. Goulash provided logs of user interactions captured across the app: signals indicating engagement with or disinterest in specific dishes, including adding a meal plan to favorites, selecting a plan that contained a given dish, or actively removing a dish from the plan, and other engagement signals. This data reflected actual human behavior, providing a rich complementary basis for personalization.

A pipeline built to grow

We began by creating a simple workflow in Amazon SageMaker Pipelines — a service that allows for straightforward data flow management for training machine learning models in Amazon SageMaker and beyond. To train the new engine, Chaos Gears used the pre-built Object2Vec algorithm from Amazon SageMaker — a neural embedding architecture well suited to learning relationships between entities, in this case between users and the dishes they engage with. The model learned relationships between users and dishes from interaction data, providing an additional behavioral signal designed to be combined with Goulash's existing content-based recommendation logic. An AWS Lambda function handled data processing, keeping that stage serverless and cost-efficient.

Once the baseline pipeline proved viable, it was extended with the components that transform a working prototype into a trustworthy MLOps system:

• Hyperparameter tuning — automated optimization of model configuration using Amazon SageMaker Automatic Model Tuning, reducing the manual effort required to improve model performance,
• Model quality validation — systematic evaluation of each trained model against defined performance benchmarks before it advances in the pipeline,
• Quality gates — checkpoints that halt progression if outputs fall below acceptable thresholds, ensuring no low-quality model reaches the next stage,
• Report generation — structured output in multiple formats, designed to be readable and actionable by engineers without data science expertise.

The modular architecture of Amazon SageMaker Pipelines was key to this expansion. Adding each new capability required no redesign of what already existed — a practical demonstration of why the choice of infrastructure matters as much as the choice of algorithm, because that flexibility allows the platform to evolve alongside the product.

The pipeline Chaos Gears delivered gives the customer something more valuable than a single trained model: a repeatable, auditable process for running experiments, validating results, and making informed decisions about where to invest next in personalization.

At every stage, the architecture was designed with security, scalability, and regulatory compliance in mind, ensuring the resulting engine could evolve into a production capability without requiring fundamental redesign.

Outcome From prototype to production-ready MLOps

By partnering with Chaos Gears, Goulash validated the business potential of adding a behavioral learning engine alongside its existing recommender before committing to a full-scale production investment — exactly the kind of risk-managed decision-making that separates companies that scale sustainably from those that over-engineer.

The project established a clear foundation for the next stage of product development and along the way, we equipped the client's team with the insight and practical knowledge to take that next step with confidence:

• Validating the approach — the client gained concrete evidence of whether a behavioral model could meaningfully improve the personalization of recipe recommendations, replacing assumption with measured signal,
• Controlling risk — by executing within a defined scope, Goulash minimized both business and technological risk and made informed investment decisions before scaling,
• Organizational readiness — the project prepared the team and the product architecture for the next phase of machine learning feature development, closing the gap between data capability and product execution,
• Retention and engagement groundwork — the pipeline establishes the infrastructure needed to drive measurable improvements in how long users stay and how often they return, through progressively more accurate recommendations,
• Accelerated iteration — by automating the MLOps workflow on Amazon SageMaker Pipelines, the time required for subsequent experiments and model iterations is significantly reduced, lowering the cost,
• Full data ownership — the solution runs entirely within the client's AWS environment, ensuring full control over data and compliance with European data processing regulations.

One recommendation engine, many pieces

Strong, market-standard recommendation systems are never a single algorithm — they are composed of complementary pieces, each solving the specific problem it was designed for. Here the two pieces divide the work naturally: the content-based system is an excellent cold-start solution that learns from the preferences users state, recommending well from a user's very first session; the machine-learning engine is built for what only accumulated behavior can reveal — taste trends across the user base and changes in each user's preferences over time. Goulash's content-based system remains the core of the recommendation feature and continues to power recommendations in the live product. The project delivered the behavioral component and a repeatable MLOps pipeline, with selected parts integrated into Goulash's production environment — ready to add its signal as user interaction data accumulates to the scale where behavioral learning delivers its full value, and giving the team the ability to evaluate, improve and expand the approach incrementally. The two approaches are built to work together, not to succeed one another. In this project, each party contributed its piece of the puzzle — Goulash the proven content-based core, Chaos Gears the behavioral learning capability and the pipeline to operate it — raising what the overall engine can deliver.

The project gave Goulash what mattered most: evidence

The result is not a replacement for Goulash's recommendation system — it is something more strategically valuable: a validated behavioral building block and a solid, reproducible foundation from which Goulash can make informed decisions about where to invest next in product personalization, machine learning and AI capabilities, with the confidence that those decisions are grounded in real evidence rather than educated guesses.

Successful AI projects rarely begin with large models. They begin with the right data and measurable hypotheses. By validating personalization enhancements before scaling them, Goulash reduced implementation risk, preserved engineering focus, and established the data foundation required for future capabilities. For organizations building data-driven products, this approach consistently produces better outcomes: validate first, invest second. That is exactly what Goulash chose to do, and what Chaos Gears helped make possible.