Pantricks

Inspiration

Pantrix began with a realization that completely changed how we thought about inventory management. Most kitchens are not bad at tracking inventory. They are bad at acting on it at the right moment. Data already exists everywhere, in spreadsheets, POS systems, and even handwritten notes taped to fridges. Yet stockouts still happen during rush hours, and food still gets thrown away at the end of the week. The problem was not missing data. It was missing decisions.

That insight inspired us to build Pantrix, a copilot for inventory management that does more than show numbers. We wanted a system that actually operates on inventory data and tells kitchens what to do before problems happen. Not another dashboard, but something closer to an autopilot.

What it does

Pantrix turns raw inventory data into real-time decisions. Instead of asking managers to interpret charts or manually forecast needs, Pantrix continuously monitors inventory and predicts what is likely to go wrong next.

Every five minutes, the system evaluates current inventory levels and forecasts two things over the next 72 hours: the probability of stockouts and the amount of potential food waste. These predictions are immediately translated into alerts that the system can act on.

On top of those alerts, Pantrix uses an AI agent to generate concrete next steps such as draft purchase orders, prep adjustments, or inventory tasks. Managers stay in control through approvals, but the heavy thinking happens automatically. Inventory management shifts from reactive problem-solving to proactive planning.

How we built it

Pantrix is built around a prediction layer and an action layer.

At the prediction layer, we trained two machine learning models. A classifier estimates the probability that an ingredient will stock out within the next 72 hours, and a regression model estimates how much inventory is likely to be wasted if no action is taken.

Inference runs every five minutes, and the results are written directly into an alerts table connected to the web application. This makes predictions part of the live system state, not static reports. To simplify deployment and reliability, we hosted our models using Hugging Face, allowing the application to call them without managing complex infrastructure.

On top of this, we built a Gemini-powered agent that reads the alerts and converts them into operational steps. Instead of surfacing raw predictions, the agent translates them into actions while keeping humans in the loop for approvals. Additionally, we used Gemini to assist intended users in understanding statistics and graphical representations effectively.

Challenges we ran into

The biggest challenge was not building the machine learning models. It was integrating them into a real-time web system that needed to be reliable, consistent, and fast.

Running inference every five minutes required careful scheduling and testing. Small timing issues could lead to stale alerts or missed updates, so we iterated heavily on our cron jobs and data pipelines. Another challenge was maintaining strict feature consistency between training and inference. Even minor mismatches could break predictions in subtle ways.

We also had to design around trust. Predictions alone are not enough if users do not believe them. That forced us to think deeply about transparency, human approval flows, and how AI recommendations are presented.

Accomplishments that we're proud of

We are proud that Pantrix is not just a demo model or a static dashboard. It is a fully operational system that connects machine learning, real-time inference, and AI-driven decision-making into one workflow.

We successfully built an end-to-end pipeline that runs continuously, updates predictions in real time, and converts those predictions into actionable steps. Most importantly, we created a system that feels practical and usable, not experimental.

What we learned

This project taught us that building useful AI systems is as much about engineering and design as it is about machine learning.

We learned the importance of separating training data from operational state, maintaining strict feature consistency, and designing systems that users can trust. We also learned that AI is most powerful when it reduces cognitive load instead of adding complexity.

Pantrix reinforced the idea that the value of AI is not in predictions alone, but in what those predictions enable people to do.

What's next for Pantricks

Next, we want to expand Pantrix beyond alerts into deeper operational intelligence. This includes better demand forecasting, supplier-aware ordering, and tighter integration with POS systems.

Long term, we see Pantrix becoming a true inventory autopilot for kitchens, one that continuously learns, adapts, and helps teams operate with confidence instead of guesswork.

Inventory does not just get tracked. It gets intelligently managed.