Update for the week ending on Friday, Jun 21, 2019

Zestful

• Added a confidence field to the API, so now Zestful returns a score of how confident it is in the result
  • This has been helpful for me in investigating inputs where Zestful gets the right answer but has low confidence, which usually indicates low quality training data I need to clean up.
• Had a call with a customer who’s using Zestful in the data pipeline for their mobile app
• Switched to a C-based YAML parser for a 6x speedup in parsing speed
  • Now I can load the training corpus in 20s rather than 2 mins
  • Sidenote: Getting the yaml.CLoader to work is surprisingly complicated
• Lots of cleanup on training data format
  • Converted all of my legacy training data from CSV to YAML
    • YAML is a much nicer format for recording the label for each token in a string
  • Got rid of lots of null values in my training data
  • Got rid of all the places that were writing labels with inconsistent capitalization
• Added logic to more aggressively pre-process ingredients to throw out tokens I know will be garbage
  • e.g., “see note”, “for garnish”, “to taste”
  • And then fixed a crash because a customer discovered that I wasn’t correctly handling the case where the ingredient is all garbage (whoops!).
• Added about 500 new labeled examples and pruned another few hundred bad legacy examples out of the corpus
• Removed about 400 lines of code that were just adding noise to the ML training pipeline
• Refactored the interface with my CRF engine so that I can take another stab at swapping out the implementation with a Python-native one
• Reached out to a potential customer who manages a WordPress Recipe plugin
• Reached out to two potential customers from Github

New project research

No progress this week! I was focused entirely on Zestful.

Is It Keto

• Joined Ezoic as an ad network and started running display ads, not sure if I’ll keep them
  • Desktop before and after
  • Mobile before and after
  • I feel like it makes the site look incredibly spammy
  • It adds ~$7 RPM (They give me about $7 for every 1,000 visitors) to the site revenue
  • Revenue prior to that was ~$9.70 RPM from Amazon Affiliate program, so it adds about 40% to my revenue, but like 200% to my shame.
  • Current plan: Try it out for another week and dump it if it still makes the site look as crappy once their ad placement tuning levels out

mtlynch.io

• Sent draft of next post to my editor
• Worked with @LoLo_ology to finish the cover image for my next post

Misc

• Tested out SaaS Pegasus and gave feedback to @czue (I think I’m the first paying customer)
• Worked with @dtq and @JDmasco2187 on hosting the first ever Western Massachusetts Indie Hackers Meetup
• Attended the Massachusetts Beekeepers Association’s 2019 Field Day