Kirby Nitzschke

Who am I

Hi there! I'm a Staff Data Scientist who has significant experience building, mainly, NLP, AI systems for production products.

Right now, I work at a customer experience analytics company where I work in the DS department. I have been at the company almost 9 years now and have held multiple different positions within different departments.

Throughout my time in DS I have been able to work on many things including:

Financial + CX linkage ML models
Training and putting into production encoder-based (BERT) and fine-tuning decoder-based (Llama, Mistral) models for predictive metadata tagging in data ingestion
Production conversational multi-agent systems as "data analysis assistants"

My path into AI wasn't a straight line. I went to school for marketing research and found my first role in traditional research approaches, survey theory, and statistical data analysis.

Through my time as a researcher (in the traditional data scientist/data analyst sense) I got many great opportunities to hone my skills of "finding the insights" in data, creating a result narrative that made sense and drove to clear actions, and present those results to many internal and external, including VP and C-Suite level, client stakeholders.

However, I was always amazed at how cool the data scientists were at the company, so when I saw a chance to join them I leveraged my way onto the team and haven't looked back! I went from someone who analyzed customer feedback to someone who builds systems that can reason about it autonomously.

What I'm Into

At work, I live at the intersection of AI engineering and data infrastructure in my day to day. I have tried to assemble skills A-Z across what a data scientist could touch. I truly love creating a model or a product that will have true impact on the company or make someone's life much easier. The umbrella term of data science can mean so many things, and I love that you get to learn something new under it every single day!

After hours, I have a recent addiction to F1 and all things racing (even got a beginning of a race sim!). I love sports and I love hanging with mates talking and watching sports. Outside of TV adjacent activities, I am usually exploring new breweries, window shopping with my beautiful fiancee, playing with my dog Harry (Sir Harold is his gov't name), or golfing.

On the learning side, I try to learn as much as I can after work - whether that is a new blog to read, a new paper that I see on X, or listening to a recent episode of Latent Space or something similar. Bridging the gap between academics or other industry examples back to my day to day job is always a fun challenge I embrace. Ask me what I have been up to lately - I keep a running tracker of everything I've read — partly to stay sharp, partly because the field moves so fast that last month's breakthrough is already old news.

Experience

Data Scientist; Senior Data Scientist; Staff Data Scientist

SMG - Service Management Group | May 2022 - Present

Architected production multi-agent QA system using LangChain and ReAct framework with supervisor-worker orchestration pattern, delivering 10 net-new capabilities (multi-subject search, NL filtering, dynamic tool orchestration). Reduced feature development time by 75% through modular agent architecture redesign, enabling new analysis tools to ship in less than 1 week vs. previous 1-month cycles. Implemented intelligent query caching system reducing follow-up query latency by ~99% (from 60+ seconds to sub-second response times)
Developed and fine-tuned generative LLM model which led to 3.5 times more insights extracted per comment, 5 ppts increase in average F1 performance across all tasks, and 98% reduction in training data.
Increased ML inference throughput by 630% while reducing annual costs by 12-29% through systematic evaluation of 4+ quantization methods, async code redesign, redis caching, and Triton hyperparameter optimization.
Enhanced model efficiency within 6 months by implementing a self-developed active learning technique to train custom client auxiliary category models, significantly improving deployment speed and decreasing needed hand labeled data.
Trained industry level models using non-linear spline regression and derivative analysis to identify most impact features to customer ROI. Wrote inference and hosting (FastAPI) codebase for API.
Won Most Impactful to Clients and Best Intentions Hackathon awards for adding intent classifier to existing production XLM-RoBERTa model.

Research Analyst; Senior Research Analyst 1&2; Senior Research Scientist

SMG - Service Management Group | Feb 2017 - June 2022

Served as team leader of 8 research analysts, facilitating daily standups and mentoring junior researchers
Developed custom ETL and built internal Streamlit dashboard for SMG Benchmarking product which increased client deliverable benchmarks by 5 times and decreased monthly labor by 25%. Won SMG Innovation award for the dashboard.
Leveraged deep knowledge of research methods to provide insights into the customer and employee satisfaction ratings to drive business results and ROI for both McDonald’s and Chick-fil-A as sole researcher on team - 2 of company's biggest clients.
Created automated weekly reports for COVID monitoring. Identified PPE desire before officially mandated, and visualized PPE performance “hot spots” in greater need of PPE. Results allowed McDonald’s to equip stores before mandates and supply chain issues, updated companywide employee policies, and decreased customers citing PPE concerns by 16 points.
Identified max time and distance for regional pizza client's delivery area. Insights led to client action which resulted in double digit percentage point increase in "orders delivered on promised time" and 3 ppts YOY comp sales.
Used decision trees to identify main impacts on store level skipped visit averages for Chick-Fil-A customers.
Used RFE and linear regression to predict comp sales for a national Retail brand. Results informed client on "in store” and “out of store” variance, which guided their CX planning sessions.
Created python scripts for a transaction data cleaning pipeline to drive efficiency and decreased labor hours by 80% for regional grocery chain.
Analyzed customer experience and delivery timer data for regional pizza concept. Was able to identify tipping points in delivery satisfaction by time of execution and distance to deliver. Analysis results and client's actions increased delivery promise time execution by 13 percentage points, increased number of orders fulfilled under 30 minutes by 14 points and increase comp sales by 3 points.

What This Site Is