# Kirstin P.

> Software Engineer at Groq

Location: San Francisco Bay Area, United States
Profile: https://flows.cv/kirstin

Exceptional developer with a strong background in math, and a wide range of experience designing and building scalable web applications, performing numerical analysis and creating algorithms, designing data models and information systems.

I am interested in data science, typed functional languages, complex networks, algorithms, service architecture, analytics. In general, all areas where math and computation intersect. Specifically, problems which require original thinking. 

Languages: Haskell, C/C++, Scala, R, Python, Ruby, Lisp, Agda, SQL, Flavor of the month.

## Work Experience
### Software Engineer @ Groq
Jan 2019 – Present | United States

### Data Tech Lead @ IMVU
Jan 2015 – Jan 2019 | San Francisco Bay Area
Introduced concepts of product similarity and product compliment based on outfits using the  appropriate machine learning algorithm, giving IMVU to have the ability to recommend  "This hat is like these hats", "This hat goes with these shoes", and "Here are dresses that go with this hat and those shoes" covering 50% of 25 million products.

Introduced Spark, configured it to work over a large hive cluster, and tutored data team in scala / spark. A 10x performance boost aside from a host of other benefits.

Grew Franz, a high performance multi-threaded event ingestion application written in Haskell to ingest events and feed them to Kafka with low latency.

### (Service/Functional Language) Consultant @ Clinkle
Jan 2013 – Jan 2014
Provided guidance for the transition to typed-functional programming in Scala from an originally Java code base.

Demonstrated use of Gatling as a library rather than an application to unify load and functional testing, within arbitrary JVM frameworks.

Added instant account verification to the backend banking services.

### Data Scientist @ Mulu
Jan 2014 – Jan 2014
Designed and implemented prototype algorithm to extract brand names from product names, and  algorithms to classify product mentions in text selections, which were then implemented in the production application. These help cut down processing time as well as help eliminate false positives.

Emphasized importance of ground truth as a model metric, and used MTurk to generate ground truth in order to test algorithm / model improvement iterations.

Constructed a tweet-music metadata matching algorithm. It used the music corpus to identify music stop words and music word importance, a tweet parser which extracted and grouped important tweet words, and a relatively small number of parameters so that it wouldn't require a great deal of ground truth to train. It could eliminate almost all false positives while generating only a small number of false negatives.

### Senior Software Engineer @ Kaggle
Jan 2013 – Jan 2013
Implemented a monadic F# evaluator for typed dataframes allowing user specified field and row transformations. Multiple user transformations were fused together so the dataframe modification was performed in a single pass.

Invented a robust (that is, it worked well on a wide variety of distributions--not just normal distributions) extreme value anomaly detector algorithm.

Wrote a probabilistic column type detector which worked reliably with only a small sample of the data, even with the presence of out of bound values.

Extended the secure Azure blob to Amazon EBS file migration infrastructure to work in both directions as well as support both user- and project-centric files via typed queue messages.

### Senior Software Developer @ Mulu
Jan 2012 – Jan 2012 | Palo Alto
Increased the speed of front page loading from over 7 seconds to under a half second through a by-need, generational, asynchronous-queue stream regeneration and an in-memory key-value store. This supported A/B testing of multiple stream generation algorithms to improve the user response rate, of which I designed several—social as well as active- and passive-behavior based using positive and negative knowledege.

Researched machine learning algorithms to cluster similar products, users, and to intelligently match users with the products they would be more likely to purchase, primarily focusing on efficient, distributed latent semantic analysis.

Audited database to ensure referential integrity and eliminate bad data, while improving the response speed.

Contributed new ideas and pushed for themes such as fun, gamification, and a conversational interface metaphor with engaging, intelligent responses to user actions, many of which were adopted.

### Consultant @ Helicor, Inc.
Jan 2008 – Jan 2012
Re-engineered burning research builds on production units, sysadmin stuff.

### Senior Developer @ Helicor, Inc.
Jan 2003 – Jan 2008
Awarded patent for work designing very low lag sparse adaptive noise filter and spectrum analysis algorithms, used on the StressEraser, a hand-held, FDA-controlled, biofeedback device for therapeutic stress reduction. Implemented algorithms and graphic user interface for this embedded device in C with 64k RAM, 8 bit CPU, and no floating point math. (C, Haskell, cyprus, ARM)

Configured and deployed secure, high-availability debian linux production servers, and built a custom e-commerce website with a content management system, discount and campaign management system, user administration with pervasive role based access, and a sophisticated automatic order processing system, with remote xml calls to the merchant bank and fulfillment center and a SOAP api for call center interaction.  (Ruby on Rails, MySQL, XML, Linux, dspam, apache2, dovecot, exim4, heartbeat, lvm, raid, drbd, shorewall)

### Senior Software Developer @ Wireless Generation
Jan 2011 – Jan 2012
Helped add support for DIBELS Next and grades 4–6 to Burst, a skills-based, assessment-driven, early-learning-literacy curriculum and instruction. The students are grouped into small, tightly focused groups by a sophisticated algorithm which optimizes the most needed skills instruction for each group. In the process of this monumental 18-month task, I led the effort to move the burst assessment back-end to a service oriented architecture.

In the process, I invented a symmetric, layered, cellular service architecture in place of MVC which emphasizes business logic in the service layer rather than MVC's focus on the interface and persistence mechanisms, and which was adopted by the team as the service architecture for all parts of the Burst application.

I spearheaded an effort reduce the size of the code while improving the clarity, simplicity, and reliability through the introduction of "traits"; the use of pure, richly typed model objects in place of integer ids and strings; insisting on semantically meaningful expressions by building up a DSL rather than dense micro-coding; and by writing a higher-order functional collections library--which led the development of a style of operating on deeply nested collations (rose trees) that greatly simplified the transformations of the highly complex data.

During this effort, I discovered that the introduction of multiple licenses and instruction sets produced incorrect historical data. Without direct support in the legacy model or data, I developed a sophisticated algorithm for inferring the correct institution and license at a given date in the past, and demonstrated the need for a dedicated licensing service (along side our assessment, instruction, authentication, and enrollment services)

Finally, I gained experience being a bettor mentor to more junior developers.

### Consultant @ Peek
Jan 2010 – Jan 2011 | New York
Called in to help project in serious danger of not meeting the delivery deadline: porting the Peek embedded client software to Qualcomm phones for BYD. Designed and built a persistant data model with a simple API to support push delivery of email, twitter, text, and facebook messages on inexpensive feature phones, which was implemented using a higher order functional collections library running on top of a double linked list with very tight memory constraints.

### Consultant @ Lumé Camera
Jan 2010 – Jan 2010
Did research and development for a pre-startup, Lumé Camera, which involved patent research as well as developing a new image processing algorithm to be embedded in the firmware. Used Haskell (a functional language--one might have heard of Scala, which is roughly similar) to design and develop the algorithm, with the intention of eventually porting it to C/C++ and OpenCL (which is an API for general purpose parallel programming on GPUs).

### CTO @ Markinekt
Jan 2008 – Jan 2010
Primarily sysadmin stuff. Migrated and consolidated services from aging RedHat installs to clean Debian and Ubuntu Xen virtual boxes. Network routing, QOS design and implementation, firewall configuration, ISP migration, VOIP routing and configuration. Also put together a working single-sign-on OpenLDAP/Kerberos/NFS4 unified system prototype.

### Consultant @ Indiana Productions
Jan 2002 – Jan 2003
Recovered and secured server taken over by hackers. Analyzed logs to determine breech. Built a custom server with grsec patched kernel, access control lists, intrusion detection (AIDE, snoop, logcheck). Installed secure, jailed services. Monitored and fine-tuned system over use. (Linux, Apache2, dovecot, exim4, lvm, raid1)

### Consultant @ Primedia
Jan 2001 – Jan 2002
Built the dynamic portions of a bridal magazine site feature (ASP, Transact SQL, SQL Server, NT, Linux)

### Consultant @ Webclipping, Inc.
Jan 2001 – Jan 2001
Wrote web robot optimized for speed and scalability. It featured pipelined bursts through HTTP/1.1, a persistent shared memory cache, and hybrid parallelization via posix threaded queues and <code>poll()</code>, which worked portably across interprocess communication based on <code>mmap()</code> or SYSV IPC operating systems. (C++, Solaris, libwww, Linux)

### Information Systems Designer @ Sticky Networks, Inc.
Jan 1999 – Jan 2000
Designed and helped implement a strongly-typed semantic network database engine for dynamic websites. It allowed multiple soft hierarchies to be generated from the same dataset (different views of the same data), could perform operations on information coodinates (contexts, like queries), so, for example, there was no difference between browsing and searching. It prefigured many of the ideas behind the Semantic Web in database form. (Java, BEA WebLogic, Oracle, EJB, Together/J, ER/win, Solaris)

### Lead Engineer @ Waysecure, Inc.
Jan 1998 – Jan 1999
Designed and implemented object database for real-time security analysis. Created template library that allowed bidirectional entity relationships for arbitrary cardinalities.  (C++, ObjectStore, Together/C++, NT)

### Senior Software Enigineer @ Earthweb, Inc.
Jan 1996 – Jan 1998
Developed new architecture for Developer.com, migrated it to Oracle, and added search.

Architect for Earthweb's largest project, USWest's DiveIn, a complete Web-Based content publishing, editorial, administrative, and statistics system with a flexible and innovative categorization system. Designed schema and implemented entity layer in stored procedures. Creative Consultant on AOL's Digital Cities project.  (SQL, ER/win, Informix, Solaris, Irix, server-side JavaScript)

### Software Engineer @ Sanford C. Bernstein
Jan 1994 – Jan 1996
Presentation creation/filing system. OO wrapper for Acrobat API and plugin for automatic page numbering, free form text search search engine for relational data, and a simple document management system.  (C++, NT, SQL, VB, Er/win, Sybase)

### Research Assistant @ Columbia Astrophysics Laboratory
Jan 1989 – Jan 1991
Developed complex modeling (analytic and Monte Carlo) and data x-ray detector placed on a Russian satellite. Wrote research paper.  (FORTRAN, Convex Unix)


## Education
### Graduate work in Physics
Columbia University

### BA in Physics, Math
Columbia University


## Contact & Social
- LinkedIn: https://linkedin.com/in/kirstinrhys

---
Source: https://flows.cv/kirstin
JSON Resume: https://flows.cv/kirstin/resume.json
Last updated: 2026-04-11