AgStacked helps businesses reduce food waste by predicting quality in the food supply chain.
Highlights
- Predictive risk profiles: Built ML models that predict container quality across 30+ produce types, achieving 80-90% accuracy and correlating with reduced claims
- Full-stack platform: Developed an end-to-end dashboard (Laravel, React, Inertia) with alerting, scaled on AWS, praised by users for making complex supply chain data easy to understand
- Integrations framework: Built a data ingestion pipeline connecting ERPs, APIs, and file transfers to unify fragmented supply chain data
- LLM-powered extraction: Initialized the AI framework that leverages LLMs to extract structured data from PDFs, packing lists, and Excel sheets
- Founder: Spoke at conferences in Berlin and the US, led sales, go-to-market, and continuous user research alongside full-stack development
AgStacked came at a time when the previous startup had finished and I was diving into a completely new domain: agri-food supply chain. Diving into a new domain is never easy, and agriculture is a particularly deep one. The early challenge was identifying what to build, for whom, and what problems were actually worth solving. You learn very quickly that you need to start with a niche.
Finding the problem
We started by building prototypes in Figma and using them to ask the right questions to retailers and sourcing companies. We found the problem: the supply chain was fundamentally reactive. Importers would open a container and hope for the best, not knowing what to expect. We also identified the size of the problem (how many companies were struggling, how much was being wasted) and what already existed in the market. There was a unique opportunity to focus on the entire supply chain rather than a single point in time.
Building the platform
Data quality was the foundation. We built an integrations framework that could connect to all kinds of platforms (ERPs, APIs, file transfers) and ingest data into a central system. On top of that, I built a full-stack application using Laravel, React, and Inertia that gave users a dashboard with alerting features, scaled on AWS. But the real question was: what do we actually show users to help them become proactive?
Predictions
We developed a prediction technique that builds what we call a risk profile over time for a particular container. As more data becomes known through the lifecycle (from pre-harvest to sensor data, cold chain conditions, delays), the risk profile becomes more accurate across 30+ produce types. When a prediction turns red, the system sends explainable notifications to end users, allowing them to assess and prioritize containers ahead of time rather than reactively opening them.
Unlocking trapped data
A large chunk of supply chain data still lives in outdated formats (Excel sheets, PDFs, packing lists) that aren’t available through traditional techniques. I initialized an AI framework leveraging LLMs to extract structured information from these unstructured formats, unlocking data that was previously impossible to access for downstream processing.
The result
Our predictions achieved 80-90% accuracy and we noticed a correlation with reduced claims for the businesses using the platform. We gave large retailers in the UK complete visibility into their supply chain for over two years. Users consistently praised the interface for conveying complex situations smoothly, and the recommendations made them want even more out of it. They could finally be proactive about what to do with their containers.