Jonathan Calvert

At Taxi Magic a core product was our mobile app that would allow a user to request and pay for a taxi, similar to other ride apps like Uber or Lyft. Drivers did not always verify they had the correct person in the vehicle and sometimes users would get in the wrong cab.

To address this, we needed to assign a confidence value that the rider and driver had successfully met so that we could escalate the issue and be triaged by customer service, and I was asked to come up with a solution.

We had GPS data from a device installed in the taxi and also GPS data from the user's device all stored in a PostgreSQL database. I implemented a Ruby on Rails based microservice that had a single endpoint that would take the user's ID, a vehicle's ID, and a time range. The service would query GPS data for both the user and the vehicle in that time range and using the Frechet distance algorithm, calculate the area between the two-dimensional linear representations of the reported route. From there, we could then assign a percentage chance that the two were together. Our main app would trigger a job to query this service a few minutes after a driver registered a pickup, and that job would then start workflows based on the result.

At Vetstreet, one of the primary services it was able to provide were highly targeted marketing campaigns acting as an intermediary for pharmaceutical companies while keeping customer data confidentiality. These campaigns would require multiple variations in the offer and language used depending on pet demographics (i.e., owners with two or more bulldogs over the age of seven). In addition, we needed to be able to service requests from multiple distinct business units.

I designed a singular Java web service that provided both a REST and SOAP interface for consumption by our analysts. Our mail provider ExactTarget required dozens of API calls per address to upload customer data across dimensions and multiple steps which meant significant time cost. In order to provide maximum parallelization, I used the Akka library which provides an actor model of programming. Each step in the creation of a campaign and data upload could then be encapsulated in its own threaded process across multiple servers, and being aware of the dependencies, it could properly cooperate to correct the order and recover in the case of failed API calls. This allowed us to scale up to deliver millions of customized, trackable product rebates per campaign.

While working for Congressional Quarterly, the company had a great deal of incoming data regarding proposed bills, amendments, regulations, and editorial material. With such a deluge of material, one key aspect of legislative intelligence is being able to determine which particular items are relevant to a particular domain as many pieces of legislation impact industries in unintuitive ways.

I designed and built a system that provided the automatic tagging of documents as a means of both broadly and narrowly classifying their topical relevance. All of our documents were ultimately fed into a MarkLogic XML database which had a means of doing entity extraction—a way of detecting things in document text such as proper nouns (people, places, products). Using this entity extraction, the code was then able to upload entities to Apache Mahout, a Hadoop machine learning library. This would apply naive Bayesian filters, and the code was able to determine the relevant classifications of a given amendment or document. These classifications could then be compared against a body of thousands of documents which had been tagged previously by the editorial staff, allowing us to instantly tag documents and alert clients in areas of interest.

At Kit Check, it was necessary to provide clean demo environments to dozens of sales agents in the field who could tailor the experience without interfering with others' work, and for a variety of reasons, it was not practical to simply host and provision disparate environments. As a result, I wrote an open source library to solve this problem.

These Ruby on Rails applications rely on a library called Rack to provide abstractions between HTTP servers and Rails applications. The rack allows you to insert middleware which enables performing functions outside the usual application code within the request/response lifecycle.

The life_guard middleware looks for a custom request HTTP header and then connects the process of that request to the correct underlying database. This approach keeps tricky code out of filters and out of the main codebase. Because it is loaded as middleware, this also allows users of the library to constrain the environments in which it is loaded, keeping otherwise risky code segregated from normal production environments.