our drink dispenser!
We are vying for the Most Entertaining Hack.
We took massive inspiration from the flow chemistry movement, where chemical syntheses can be finely controlled with different reaction parameters, leading to efficient, and eco-friendly laboratory processes. Our original idea was to incorporate enabling technologies such as machine learning and open-source hardware into chemistry research.
However, due to the hardware limitations of this hackathon, we decided to explore the world of taste and aroma. We decided that with so many different cuisines and dishes, it would be a promising development to create a machine-learning assisted automatic bartender that mixes drinks according to what food you're having.
What it does
The user enters a dish as an input, and the web app we've designed deconstructs the dish into its key ingredients, scraped from the Yummly API. Afterwards, the ingredients are further classified to 14 unique flavour profiles, which is translated with a single-layered neural network into a bespoke drink mix, from five drinks that the neural network feels that matches the dish best.
How we built it
Firstly, we came up with original designs to create a water pump from a DC motor. This required immense effort on our part, including making 3D printed parts and interfacing with an Arduino.
The front-end web app was done up in HTML/CSS, with a web server and database in Node.js. The neural network was created with TensorFlow.
Accomplishments that we're proud of
We're proud that we've managed to create a single-layer neural network, with 20 hidden notes, to pair 14 different taste profiles from various cuisines into a mixture of five different drinks.
In addition, we made a prototype of our drink dispenser which demonstrates the efficacy and feasibility of our installation. We are extremely proud that we our dispenser with recycled materials such as cupcake boxes and plastic parts we found in the recycling bin.
Challenges we ran into
Due to time constraints in training, we were unable to move beyond our single-layer neural network. In addition, we did not have hydraulics or fluidics equipment on us to manipulate liquids, and thus we faced numerous process engineering problems such as pressure loss due to non-watertight components, which led to problems with pump operation and implementing electronics.
What we learned
We learnt that developing hardware hacks is difficult, and requires a substantial inventory and familiarity with tools, so that time is not wasted sourcing for and optimising ineffective parts.
What's next for TensorOverflow
We intend to implement continuous online machine learning in future iterations of the project. In addition, we're very interested in pursuing fluidics processing to analyse, optimise and exert fine control over chemical reactions and create useful tools for laboratories in the future.