-
Going to try to pick up Elm again, for the purpose of writing a game. Tried last year, let’s see if it goes better this time…
-
-
For school, we have to think about entrepreneur projects related to blockchain. Every single time we find a nice idea, either it has already be done, or the blockchain technology is irrelevant for that idea (which can be done without). Our group hasn’t advanced for 2 months now.
-
-
-
-
-
I wrote an article on finding ideas. Essentially, it is important to find problems and treat them as opportunities, rather than finding solutions first and the problems they solve later.
-
-
-
-
Maybe start from problems: look for markets for lemons, adverse selection, agency costs. As a rough rule of thumb, any market in which someone can earn a commission. And focus really tightly - rather than land titles on the blockchain, attack mineral or oil rights. Look up what people are suing each other over and you know what corner cases to handle.
-
Things that might be useful:
- Many companies start from open-source projects aiming a niche, enhance them a bit/lot so it can be useful for companies.
- Others are just something that already exist but can be better/interesting on the blockchain. For example, several softwares to track authenticity of painting/sculpture transactions exist, but blockchain ones emerged and it’s a super neat application (imho).
- You probably do this already, but if you shift your view of the blockchain to either tokens or just a more realistic representation of the world, it might be easier.
-
-
That’s actually a good point. YC often says don’t worry if someone has thought of your idea already. Just beat them in execution. Tech history is littered with better ideas or bad implementations of similar ones that lost to better executed and/or marketed ideas.
-
-
Although I warn it might be unpopular, you might want to try something similar in concept but not quite blockchain. The benefits of the blockchain without necessarily being one. Here’s a few I’ve heard or was pushing that may or may not be implemented by a startup by now:
-
Transactions are done with traditional databases that use a distributed ledger to tally up final results. This is similar to what banks already do where most transactions are hidden in their databases with some big chunks of money moved between banks. It works.
-
Instead of just a coin in the ether, Clive Robinson on Schneier’s blog suggested creating a financial instrument that is tied to a number of commodities or other currencies in such a way that it remains really stable. As in, not a speculator’s tool like Bitcoin. I found one company that did this with several currencies plus carbon credits. I just can’t remember name.
-
Instead of miners, might again use a low-cost technology for transactions but people need an account with the service to participate that costs about a dollar or so a month (or yearly). Kind of like with texts, they buy blocks of transactions. The providers are non-profits with chartered protections with the provisions or exchange being where the new tech comes in to provide accountability.
I’d do a combination of these if I entered the market. I’m not planning to right now. So, sharing the ideas with others in case someone wants to have a try at it while money is raining from the sky on those that utter the words “blockchain” or “decentralized.” ;)
-
-
-
Finally launched Helmspoint. It deploys your Keras image machine learning model to the web.
Had some hiccups along the way, but I finally have something that people can use, and I can start iterating from here.
I’m going to:
- Write down a list of people to swing back to so I can talk to them.
- Solicit feedback from new users
- Come up with a pricing plan and implement subscriptions
-
Talking to customers. Finding more ways to talk to customers. And building things at the same time. In the past, i’ve Always found it hard to do both at the same time.
-
Helmspoint, deploys your keras machine learning models.
Last week:
- Got ingress working (finally) Had to turn off the default GCE ingress that was fighting the nginx-ingress
- Finished the basic pipeline editor when creating a new version.
This week:
- Get subdomains working, so when you publish a model, it has a subdomain associated with it.
- Add TLS to subdomains.
-
Working on Helmspoint, it deploys your machine learning models to the web. Last week:
- Got resnet50 working, but ran into things I didn’t consider, like custom input preprocessing of images and prediction decoding. I ended up having to build a simple pipeline, to handle these cases.
- Didn’t get around got the nginx-ingress. This week:
- Will need to add UI for pipeline editing when creating a new build version.
- Want to generate hashes all components that go into a build version instead of using database id numbers. Not quite necessary for launch, but it’s something easier to fix now than later. Maybe it’s a form of procrastination to work on this.
- Need to get nginx-ingress up and running.
-
Helmspoint a tool to deploy machine learning models to the web
- I’m working on getting ResNet50 up and running in a deployment. It should be able to classify images.
- I’ll be working on getting the nginx ingress up and running after turning off the default ingress, so I have a single load balancer.
-
Working on Helmspoint, a tool to deploy Keras ML models to the web. Wasn’t able to get to what I said I was going to do last week.
- Fixed a major bug in production that required uploading files to cloud storage first.
- Added ability to track and deploy multiple versions of your app.
This week, I’ll be:
- Adding a UI and CSS style the thing.
- Emailing people I previously talked to about Helmspoint to get some feedback
- Research and play with Ingress for load balancing
-
Got all the kubernetes and docker stuff all set up for helmspoint–a tool to deploy already trained machine learning models to the web.
This week, I’ll be
- emailing people that use keras to ask them about their deploy setup
- adding documentation on how to export from Keras
- adding basic UI flow for user to create a project and deploy the models
-
I’m working on helmspoint, a way to deploy ML apps. I’ll be working on the docker client connection to create images. Before, I was just using popen for running docker on the command line. I’ll then need to do authentication for kubernetes and docker REST APIs.
I finished a blog post about how Norway got Japan to eat salmon sushi after a decade long campaign. https://medium.com/torodex/salmon-sushi-is-not-a-japanese-invention-9189d9cd78b7
-
I’m working on Helmspoint, a Heroku for ML apps. Still wrestling with some image setup stuff. Will have to start deploying the application itself.
I’m also working on a sushi blog post, about how delicious sushi is isn’t tied to how recently the fish was hauled off the boat. https://medium.com/torodex/the-sushi-you-eat-isnt-fresh-f3478c99eb2
-
Working on www.helmspoint.com, a heroku for keras ML models.
Finished the kubernetes and docker integration. Couldn’t actually make the docker images too much smaller. Alpine didn’t compile the tensorflow pips that I needed. Found out you can stack docker images instead. You can now upload the weights and model file, and it’ll deploy to minikube.
Will be working on making the generated app more user friendly, as well as devising a UI for helmspoint itself. Hopefully, that won’t take too much longer, and I can start working on getting it into production this week.
-
“The Dip” Just as a reminder to keep pushing.
-
Working on Helmspoint, where you can deploy your ML models.
- Right now, need to make make smaller Docker images. Otherwise, the deployment would take too long.
- Need to refamiliarize myself with Kubernetes and find a way to deploy without an image on a Docker repository (if it exists).
- Lastly, need to start working on the basic UI, so people can upload their files.
-
Working on Helmspoint, which helps people deploy ML models on the web. Just upload the model and the architecture, and it takes care of provisioning the servers and the API.
Got chain-able background jobs working with map and and_then semantics. That way, I have composable jobs that can be written in isolation from each other. And I can use composable jobs to write processes that have many steps with side effects.
I’ll be working on integration with Kubernetes and figure out whether to put it on AWS or GCloud.
-
Working on Helmspoint. It’s deploys machine learning models to the web. Currently only supporting Keras. You just upload the weights and the architecture, and we take care of provisioning the servers, the web app, and the load balancing.
Within Helmspoint, I’m currently working on making chainable background jobs. In previous work, when a user makes a web request, we want to do the work in a background job, since it takes too long to execute. However, these jobs could be complex, and might change due to changing business requirements.
Taking a cue from effects managers in functional programming, I’m separating control flow into ‘and_then’ and ‘map’ semantics, from the side effects in the jobs. That way, I have a small DSL for deploying and provision servers that’s flexible and modular. The unwieldy thing about it right now is that I need to serialize both functions and their context, to make them queueable as a job. If anyone else has a suggestions and pitfalls when it comes to serializing functions, I’d be happy to hear it.
Haven’t finished that blog post from last week, but will do so this week.
-
Working on Helmspoint, a tool for deploying Keras machine learning models to the web.
- Implement chained background tasks
- Write blog post
Individual background tasks are easy to fire off, with resque, sidekiq, and its ilk. But when I’ve had background tasks that had to be chained together, it got messy. Think I finally have a solution now, where I can build tasks that are chained with AndThen and Map. It’s almost like rewriting promises and effects managers, but now they’re persistent.
Haven’t decided on a topic, but it might be “How do I get other people to give me data”?
-
It doesn’t mention Kaggle!
-
-
This was what I was working on this week. If you got feedback, I’ll be glad to hear it.
This week, I’m doing some contract work to pay the bills.
But on Helmspoint (https://www.helmspoint.com) I talked to users and got some indirect feedback. I think I missed the mark a bit, and will need to go back to the drawing board. I use to think the data pipeline stage was different than the deployment stage when using ML in production. I believe there can be a single abstraction to use for both, and will be prototyping it out.