Inspiration

Initially I had an idea to make an autofilling PDF form but after consulting, understanding the needs of the wilderness folks, I realized something much more innovative and unique. Something that motivates young 22 year old eyeballs in congressional offices into presenting material to their congressperson.

So writing an actionable personalized letter to congress is hard, it requires creativity, context and knowing what motivates your audience into action. Machine Learning makes this easier, even trivial.

What it does

In a nutshell, I generate guaranteed unique PDFs contextualized with the user of my mobile app (like their current location, photos of the park they are in, who their congress person is, the latest zeitgeist in politics) and contextualized with what words, ideas cause and stir action in the young twenties congressional aides that read letters, emails from concerned citizens.

I do this by leveraging the latest Machine Learning research code (Released in April 2019) from OpenAI called GPT-2. This is an incredibly powerful new model insofar as there is no prelabeled data, rather the model is able to figure out context ON ITS OWN. Thus giving it just a few parameters, like the user's name, their basic interests and their political values is enough to generate and convincingly hand written letter, then generated as a PDF ready to be verified with docusign's Driver's License API, enriched with photos of the local park if so desired (think of an activist in Yosemite wanting to make their PDF letter even richer with an actual photo of Yosemite inserted in the generated PDF).

I understood this as my plan because a plain auto generated letter has people just tune out, so given that the average letter only gets about two minutes of attention time by the congressional aide, then it really needs to pop. Machine Learning handles this for me by generating the most attention pulling letter, the most convincing, the most likely to stir action and pull in the attention of the congressional representative and their aides. This is ground breaking and nothing like it exists.

How I built it

I built it with ReactNative for the mobile app, a pytorch implementation of OpenAI's GPT-2 tensorflow code, google compute because I needed a GPU, python for generating a real PDF and simple web server.

Challenges I ran into

I had an emergency on Saturday so I could only devote three hours of coding to this, on Sunday morning from 9AM to 11:20AM. Besides that, the challenge of working with research machine learning code. I haven't coded seriously in past six months.

Accomplishments that I'm proud of

I built a serious MVP in a very little bit of time, with several libraries that I never used before but figured out very quickly. My ML code is running now live on a google compute machine, my real mobile app talks to that backend and it displays the real PDF that was generated.

What I learned

The incredible power of GPT-2, the ridiculous research done at OpenAI.

What's next for docusign-hackathon

Depends if there is interest in continuing, but this could be an amazing and actionable project.

Built With

  • python-pytorch-machine-learning
Share this project:
×

Updates