Hi, I'm Leah. I like hard puzzles.

Most of mine are in AI inference: getting open-weight models to run fast and give correct answers on GPUs. I work on this at AWS and in my own projects.

About Me

What I'm thinking about

  • Why the fastest benchmark result is often the wrong one.

  • How far a model can be quantized before its answers change.

  • Real-time voice: how much can happen between one spoken turn and the next.

  • Keeping my own archive of open-weight models and running them locally.

Away from the keyboard

  • Cooking (attempts were made), and learning to make cheese (many attempts will be made).
  • Daydreaming about off-grid living on an affordable remote tropical island.
  • Speaking German (badly). I studied German in Frankfurt.
  •  

Tech Stack

Inference and Models

  • Serving: vLLM, TensorRT-LLM, Triton, KServe, NVIDIA NIM

  • Optimization: NVIDIA Model Optimizer (FP8, NVFP4), same-hardware precision comparisons

  • Benchmarking: MLPerf LoadGen, AIPerf, correctness-gated concurrency sweeps

  • Speech: Magpie TTS, OmniVoice, Parakeet STT, vLLM-Omni

Kubernetes and GPUs

  • Cluster: Amazon EKS, Karpenter, Bottlerocket

  • GPUs: EC2 G and P instances (A10G, L4, L40S, RTX PRO 6000 Blackwell, H100, B200)

  • Capacity: Spot and On-Demand tiers, SOCI parallel image pull, S3 model loading

  • Networking and Observability: EFA, NCCL, Prometheus, Grafana, DCGM

Real-Time Applications

  • Services: Python, FastAPI, WebSockets, Server-Sent Events, MessagePack

  • Orchestration: Temporal, Celery

  • Data: PostgreSQL with pgvector, Redis, Amazon S3

  • Agents and Models: MCP servers, Hugging Face, Natural Language Processing (NLP) Models

Professional Highlights

Stripe

I wrote Quickstarts in seven languages that helped raise Stripe Checkout activation 25%.

AWS

I shipped the first Amazon EKS AI/ML best practices, which earned 92% customer satisfaction.

Sabre

I led the developer portal for 2,500+ APIs across 100+ teams. Daily active users rose 15%.

Avis Budget Group (ABG)

I designed shared API data models and a design guide that cut time to market 75%.