Nginx

Upon joining, I came up to speed quickly on the stack of the early version of Pull Workbench, which was very buggy but demonstrated the initial ideas and had a solid set of the latest technologies and patterns established in the codebase, providing for a solid starting point.

I was entrusted to aid our CTO in hiring several additional employees, and so I joined and conducted interviews for the first several months while working with the existing team AI + Full Stack to deliver features and solidify the system, with the aim of keeping it fully working with each merge, after playing a little catch-up to fix the early bugs that worried our business partners, giving them confidence that our team could deliver.

From there, I developed full stack features solo or by pairing with team members, and ultimately led a squad of 5 team members alongside a second squad that together comprised our engineering team.

Much of my time went into authoring complex analytics sql queries using the impressive Kysely library, a fluent, typesafe query builder that we used for our postgres and redshift databases. Given the nature of the product, we needed to make decisions on which queries could be run in real time vs. which queries and subqueries would need to be computed offline as part of a network of airflow dags.

On the ML Ops side I advocated for traceability and reproducibility / determinism of all models and artifacts, and integrated with systems that implemented that, such as Airflow to coordinate DAGs of ML training jobs and Sagemaker's metadata API, which we controlled via model lifecycle automations that produced and stored models, artifacts and metadata that were in turn consumed at runtime or in batch by our analytics stack

On the frontend, I helped us deliver an initial version of the Pattern Editor, a UI and set of APIs that users could use to put together their own patterns of interest, such as looking for certain anomalous ranges of quantities that themselves may be derived from other user-defined patterns. This entailed not only a UI that was DAG-aware but also a layer that converted the json representation of these patterns from the frontend into typesafe kyesely queries to be executed against redshift.

Interview Builder is where customers would go to define the values they wanted to find in their ideal candidate, and map those to attributes and ultimately interview questions that the intertru ai was pre-trained to assess.

My role was to work closely with the CTO to understand what was proven out on the ML side, so that we could deduce a UI that intuitively would extract the necessary inputs from the customer, while providing them with predefined templates as starting points to ease them into the process.

The application was built on react, typescript, graphql (backed by dynamodb) and amazon amplify, and I built it very quickly with simple backends so that we could iterate on the frontend, to get the experience right before investing significant time and effort into an ideal backend. This approach made iterations faster and produced less collateral damage / throwaway code as we refined the user experience.

We then added instrumentation so that we could measure the use of the feature, any bugs that might turn up, and its performance, before releasing it to production, where it was initially used internally to surface any shortcomings before customers were exposed to it.

The candidate summary page summarized a candidate's performance during multiple interview stages by presenting radar charts showing degree of fit against the values and attributes being evaluated for their position, as defined in the Interview Builder.

I built the frontend in React and Typescript, and integrated with the backend, which I partially built, which leveraged RAG and ran several Machine Learning models to produce scores and explainable AI. For example, models to break down interview transcripts into quotable fragments, evaluate relevance against configured company values, and call chatGPT APIs to obtain summaries and scores related to that content

A working prototype and specs for how heartpoints would be rewarded and exchanged and linked to "Proof of Good" that could be validated off-chain (since proof of good in this case may for example, be video evidence or other data that is too large to fit onto the chain), using a strategy of hashing the proof and storing the hash and URL of the proof's off-chain content.

New Admin Dashboard - Architected and implemented green field admin dashboard, worked with junior devs and leadership to bring about in the requested stack: Bootstrap / JQuery / Ruby / Java / MySQL / LDAP. Users used Admin Dashboard to administer Organization, Role, and other management aspects of the Explorys EPM Suite.

Dependency Injection Framework - To enable high testability, all components required dependencies to be passed in, and so eliminating factory boilerplate without adopting a complex DI system was desirable. Wrote and shared a simple runtime-reflection based DI framework with multiple projects / teams

Temporal Sequencing - Architected, implemented and tested an enhancement to our proprietary Explorys MDL (measure definition language), which gave authors who use that language the ability to express dynamic temporal conditions as part of the predicate calculus that constitutes the measure being defined in a given MDL instance.

The solution runs in Hadoop as a mapper, where the MDL is parsed and a corresponding object graph is constructed, and then control is passed to the encapsulating object of that graph, resulting in a determination about the level of adherence of the given patient to the measure defined by the MDL which is then emitted for further processing.

The challenge here was authoring the solution without knowing the history or complexity of the existing system. The challenge was overcome by applying SOLID principles and TDD principles to express assumptions about the existing external system as interfaces, getting the solution to work in a test sandbox as a standalone application, and then once that was done, writing adapters between the actual system and the interfaces that were formerly mocked out for testing. The team which owned the system seemed impressed by and curious about the solution and the new patterns / tests involved.

By taking what had been considered a relatively tough problem, and, as an engineer with no history in the measure engine, applying an approach that allowed a fully tested solution to ultimately be plugged into the larger system very cleanly, as an implementation of a clear set of interfaces along the border of that system, I exposed the team to new patterns and solution design approaches that I hoped the team might take forward and build on as the system continued to evolve.

Example Use Case: Doctor wants measure of % of his/her patients having Hospitalization with outcome class X, where either 1 week later, Rehospitalization occurs or within 2 weeks, unexpected Office Appointment occurs. Doctor expresses using natural language oriented temporal sequencing expressions within model definition language.