David Denny

Personal · Built on CMU coursework

Benchmarks each rung of a QA stack on SQuAD: Jaccard → TF-IDF → BM25 → a learned reranker (81.6% top-1 sentence accuracy) → a transformer reader (0.87 F1). Also measures how much latency you save by narrowing the reader's input, and what it costs in accuracy.

Personal · Built on CMU coursework

A small NumPy library with a scikit-learn-style API: Newton-method logistic regression, Pegasos SVM, softmax with Adam, TF-IDF, and naive Bayes. Each model is tested for parity with scikit-learn and benchmarked on public datasets.

UT Austin · Advanced Deep Learning

Implemented half-precision, 4-bit quantized, LoRA, and QLoRA variants of a large PyTorch network, cutting weight memory from 72 MB to 11 MB while keeping it trainable.

UT Austin · Advanced Deep Learning

Wrote batched generation from scratch for SmolLM2, then improved reasoning with chain-of-thought prompting, LoRA supervised fine-tuning, and a rejection-sampling fine-tuning (RFT) RL loop.

UT Austin · Advanced Deep Learning

Built a data pipeline that turns vision labels into question/answer pairs, fine-tuned a VLM on it, then reused its encoders to train a CLIP-style contrastive model for multiple-choice QA.

UT Austin · Advanced Deep Learning

Trained a patch-level autoencoder, turned it into a tokenizer with Binary Spherical Quantization, and trained an auto-regressive transformer over the tokens to generate and compress images.

Personal · Open source

Event-driven ingestion service on the public USGS earthquake feed: validated bulk upserts with an advisory-lock guard against overlapping runs, RabbitMQ fan-out to workers, Elasticsearch relevance tuning, OpenTelemetry tracing, and Kubernetes health probes.

UT Austin · Deep Learning

Trained MLP, CNN, and Transformer planners that predict driving waypoints from track and image data in SuperTuxKart, comparing how each architecture handles the same task.

UT Austin · Deep Learning

Built the full transformer architecture in PyTorch from the ground up — multi-head attention, positional embeddings, residual connections, layer norm. Trained end-to-end and benchmarked across attention configurations.

UT Austin · Machine Learning

Forecasting stock-price movement on Nvidia and Tesla. Trained and compared decision trees, random forests, and gradient-boosted models against linear-regression baselines, then wrote up methodology and results end-to-end.

CMU · Applied ML & Deployment

Trained a convolutional network on CIFAR-10 and deployed it to an Azure ML container endpoint, comparing CPU vs. GPU training cost and inference latency.

CMU · Applied ML & Deployment

Built a movie recommender with collaborative filtering and matrix factorization, reducing rating MSE from 0.80 to 0.64 and comparing sparse representations for storage and runtime.

CMU · Applied ML & Deployment

Collected public text from APIs and scraped web pages, then cleaned, tokenized, and lemmatized it into a single corpus ready for feature extraction and analysis.