BENCHMARKING

mini-transformer-gpu

A custom language model architecture built from scratch in raw PyTorch. Features custom multi-head causal attention loops, BPE tokenization, and validation on consumer GPUs.

PyTorch CUDA Transformers Python
PRODUCTION

stream-elt-lakehouse

High-concurrency streaming pipeline ingesting millions of telemetry events per minute from Apache Kafka into Snowflake with real-time anomaly detection and strict schema enforcement.

Kafka Spark Snowflake Terraform
RESEARCH

surprisal-llm-eval

Psycholinguistic predictability and surprisal modeling suite. Evaluates token probability matrices across fine-tuned LLMs against cognitive human reading benchmarks.

Python HuggingFace NLP SciPy