October 9, 2019

Soft launch + first Medium blog post

JP Hwang @hindsights_io

I'm glad to say that the web app (https://www.hindsights.io) is live, which is very exciting, although we are not taking paid subscriptions yet.

For last week or so I've been working on fine tuning the app to run at acceptable speeds, making the front end more user friendly and trying to figure out what the best way to market this thing would be.

Speaking of marketing, I have published our first blog (https://medium.com/@hindsights/patent-searching-and-analysis-sucks-but-what-if-they-didnt-5bf7f651f04b) on Medium. To start with, I wrote about the "why" of choosing this startup, and plan to post a little about everything, including my personal story, tech used in building the site, and professional journey here.

I've received some good advice here, that for SEO reasons it is preferable to have / host your own blog, but as we don't have any visitors yet I thought I'd get started on Medium, and start to cross-post on my own site slowly.

I have still got a LONG way to go, but I'm excited to be moving forward bit by bit.

  1. 2

    Nice landing page. What were your thoughts behind releasing the pricing page before public launch?

    1. 1

      Thanks James :)
      I had a few reasons for putting the pricing in. I wanted to be up front about what I thought the launch prices would be for transparency, and also to let potential visitors know that these tools would be priced reasonably (which I don't think typical patent software tools are, on both points).
      I also thought having a pricing page would let me measure bounce rates of people leaving after visiting the pricing page.

      1. 2

        Great ideas!! Those are some really good points. Have you thought of putting a 'Signup for launch' call to action form in place of "Coming Soon" so you can capture emails for those who want that specific price plan?

        1. 1

          Ah! That’s a great tip! Thanks James. I’ll add that to the to-do list.

  2. 2

    Way to Hwang!! Landing looks polished (though some more color wouldn't hurt on the home page!). Is this a Django app? Kudos on the tech stack selection, heh

    1. 1

      Thanks buddy! Yes it is a Django app. I started with Flask but Django (and its batteries-included approach) just worked better for someone like me with limited experience in web app building.

  3. 2

    This comment was deleted 6 months ago.

    1. 1

      Hi Alex!
      Thanks for the encouragement and suggestions! They’re great and I agree with you. They’re on the todo in Trello now, to be implemented shortly.

      I will reach out when I need help, I undoubtedly will! It’s been interesting learning NLP / full stack web app dev haha and being a sole founder. I couldn’t have done it without help from great resources and people online like yourself, so it’s much appreciated.

      Best
      JP

      1. 1

        This comment was deleted 6 months ago.

        1. 1

          Hi Alex,
          I've posted a blog post here. https://medium.com/@hindsights/how-i-built-a-patent-text-concept-search-analytics-app-af67085b0ac0?source=friends_link&sk=1d353f753bb33f4e9e340b5fa34d6469.
          I don't know that any of it'll be new info to you but there you go :).

          Let me know if you have any tips/comments!
          JP

          1. 1

            This comment was deleted 6 months ago.

            1. 1

              No you didn't miss it :).

              I tried a bunch of different methods including doc2vec and averaging word vectors. Neither of them were not that great for what I needed to do - which was why I went and spent a bit of time coming up with my own metrics to measure intrinsic/extrinsic performance.

              Long story short, doc2vec was bad enough that I didn't ever want to touch it again haha. Personally I don't think having a single paragraph/document vector for an entire document makes a lot of sense. You're essentially performing the prediction over such a large window (of the document) that I don't think it's really predicting anything meaningful.

              Averaging words work much better, and various topic modelling algorithms like LDA and its variants work okay also.

              Ultimately I wrote something of my own, which... I don't know if it's smart of dumb but I couldn't find anything that really suited what I wanted to do!

              1. 1

                This comment was deleted 6 months ago.

                1. 1

                  Ah that's really interesting. I definitely didn't spend a lot of time on doc2vec so maybe I could have done more on that front.

                  Your comment about calculating a median word vector and constraining the dictionary are really good points! Intuitively it sounds like it'd reduce the amount of "noise" that creeps into the overall vector that is composed, whether that is for the paragraph or the document. Is that right?

                  As to what I'm doing that's "more" than averaging word vectors, they're quite patent specific things.

                  Patent documents have different sections (title / abstract / description / claims) that serve different purposes. So I actually weigh these parts differently to calculate the overall vector, because I want the vector to represent they key concepts of the invention.

                  Like if a word appears a bunch of times in the description, but not in the claims, that's intuitively an indicator that it's not as significant a word in the context of the "invention" as defined. Also, while the "claim" is the most important part of the patent, it's much, much shorter. So if I hadn't weighed the sections differently, all of that information would be lost in the "noise" of the mammoth description, a lot of which might be like descriptions of old practices or just information that's there for legal reasons.

                  Does that make sense?

                  1. 1

                    This comment was deleted 6 months ago.