LocalRecall

Inspiration

We often lose small things around our desks. Cloud cameras and APIs can help manage and keep track of these objects, but they’re a poor fit when you want privacy and no endless video of your workspace. LocalRecall came from wanting a private memory of where enrolled objects were last seen on one bounded tabletop.

What it does

LocalRecall allows you to define zones, and create named, personal belongings, as well as analyze before and after images with a local vision model on your Spark-class box. Ai Validates movement of objects from different preset zones and explains movement of objects between them. An Ask feature has also been impalaces to act similar to the ai movement validation.

How we built it

We used Gradio Blocks for a five tabbed UI and a Python facade as the only bridge between the interface and backend logic. Labeled composites go to VLLM and Qwen3.8-27B-NVFP4 through a VLMClient, then event validation decides what may update SQLite before Ask runs. Mock mode and conformance tests allow us to develop the same contracts without restariting the model or persistent storage on every machine.

Challenges we ran into

Getting Qwen to work on the GB10 was a major challenge faced during the building process. Also being a first time hacker, high-school based team made it ever more difficult to familiarize ourselves with the tech and create systems for our code.

Accomplishments that we're proud of

Fully deploying a Docker image on the GB10 was a major accomplishment for our beginner team. We are also proud that this event served as a valuable learning experience, allowing us to gain hands-on experience working with the GB10 and AI servers.

What we learned

We learned that separating scene changes, what the AI model sees, and what the database stores makes it easier to find and solve problems. The manual BEFORE/AFTER feature was one of the most important part of our project because as allowed us to compare scenes directly. We also learned that Mock Mode, clear labels, and shared validation rules helped our team work together and avoid showing incorrect information between each other.

What's next for LocalRecall

Next, we plan to improve the demo by connecting SSH forwarding, browser webcam access, and Spark USB snapshots into one working system. After that, we will work on Watch Mode, which automatically detects scene changes, but only after the manual features work reliably.