~Mitchell Diedrich~

• Designed and built TypeScript and Python libraries extending AWS Bedrock with model comparison/evaluation, cost estimation, LLM-as-judge, and guardrail controls. Shipped with runnable example apps. Selected by engineering leadership as preferred implementation for LLM interaction and as library design reference.
• Maintained the centralized LLMOps platform managing AI interaction for ~15 teams, enforcing enterprise standards and extending Bedrock's functionality with evals, observability, and cost alerting. One of two engineers responsible for diagnosing and resolving production issues.
• Cut model deployment iteration time for ~30 data scientists from 15 minutes to under 1 minute by replacing a CI round-trip with a local notebook harness.
• Diagnosed postprocessing bug in production insurance-lapse model; F1 score recovered from 0.58 to 0.70. Delivered root-cause writeup; drove data science team's adoption of peer code review.
• Built a multi-account SageMaker platform with 3 engineers, with CI/CD pipelines and a self-service project template for one-click ML scaffolding; paired with data scientists to fit the template to their models. ~50 projects deployed via platform.
• Productionized ML models across the US insurance segment, including a lead-scoring model to help sales reps allocate effort and a BERT email classifier routing inbound mail to eliminate manual triage.

• Built a process to provide developers with obfuscated database copies, allowing work with production data while maintaining regulatory compliance.
• Automated tax document processing via Amazon Textract, saving ~100 hours each week of tax season.
• Designed a MySQL log parser that traced data lineage through nested joins and views, letting engineers find where a field originated.

• Developed Keras computer vision models to classify documents, replacing manual classification.
• Developed TensorFlow computer vision models to classify patient imaging based on disorder presence.

Gradient-boosted classifier with walk-forward time-series validation over 20 seasons of player data: AUC 0.878 across 3,389 out-of-sample predictions (2010–2024); High Risk tier calibrated at 87.9% actual decline vs. 11.9% for Low Risk. Feature engineering across player stats, age curves, injuries, workload-vs-peak, and team changes.

Predicted channel-relative video performance on a 1.15M-video corpus (92k/11.5k/11.5k train/val/test); fine-tuning the MiniLM title embedding model lifted test R² from 0.06 to 0.22. Benchmarked multimodal (title text + ResNet50 thumbnail features) and transcript topic-cluster variants; documented negative results.

Library for activation-steering style transfer on open-weight LLMs: style vectors extracted from contrastive pairs in the residual stream, injected at inference via forward hooks; bridge-style composition for style pairs with no parallel data. pip-installable with CLI.

Computer-vision web app reconstructing a 3D skeleton from single-camera golf-swing footage: pose estimation, key-event detection (address/top/impact), and 8 biomechanical metrics benchmarked against a reference swing, served with an interactive 3D viewer.