Data tool · Web app
RI2 — Rapid Insights Data Engine
A no-code workspace for cleaning, visualizing and statistically testing tabular data, with real pandas and scikit-learn running in the browser through WebAssembly.
Work
Projects and experiments
11 projects across language models, developer tools, analytics and teaching. The 6 with case studies are the ones I’d most like you to see.
Showing all 11 projects
Language models · Research
TinyLM Lab
Pretraining a 26.7M-parameter Llama-style model from random initialization on one laptop GPU, then checking what it learned against memorization, leakage and controlled ablations.
Web app · Product design
Kaku — Tiny Editor
A calm, private writing room in a browser tab: rich text, stationery-inspired themes and local export, with nothing ever saved or uploaded.
Education · Analytics
The Power BI Fellowship
A free, open-source way to learn Power BI by doing the job: checkable skill exercises, then tickets, incidents and architecture decisions at a fictional company.
Generative AI · Document intelligence
Self-Correcting Multi-Document RAG
Question answering and summarization over multiple documents, with agents that filter irrelevant context, score each answer’s factual consistency and regenerate weak answers.
Computer vision · Benchmark
Deep Learning Architectures on CIFAR-100
Fourteen architectures, from logistic regression and LeNet to EfficientNet and Vision Transformers, compared on accuracy, training time and parameter efficiency.
Results
Images were resized to 224×224, and the strongest models start from ImageNet-pretrained weights, so the top results measure transfer learning rather than training from scratch. Swin-Tiny’s 3.03% most likely reflects a failed run; the repository marks itself as experimental.
Multimodal AI · Research
Multimodal Medical Visual Question Answering
Fine-tuning LLaVA-1.5-7B with LoRA to answer questions about radiology images, evaluated with a custom token-level F1 pipeline.
Forecasting · Kaggle
GoDaddy Microbusiness Density Forecasting
Forecasting monthly microbusiness density for U.S. counties in a Kaggle competition, finishing in the top 24% of teams.
Details
- Objective
- Forecast microbusinesses per 100 adults, by county and month, from GoDaddy’s Venture Forward data.
- Features
- County history plus Census ACS indicators: broadband access, college education, foreign-born population, IT workforce and median income.
- Approach
- Tree-based regressors compared on the same features; the write-up covers why Extra Trees stood out.
- Metric
- SMAPE, the competition’s official metric.
- Outcome
- Top 24% of teams on the leaderboard.
Sports analytics · Deep learning
Soccer Analytics: La Liga
A PyTorch classifier that predicts La Liga match results (win, draw or loss) from FBref match statistics, with a Streamlit dashboard for exploring team performance.
Details
- Data
- La Liga match statistics from FBref: 2015–2021 seasons for training, 2022–23 for testing.
- Features
- Goals, expected goals (xG and xGA), possession, shots, and shot- and goal-creating actions.
- Model
- A multi-class neural network in PyTorch on min–max normalized features, predicting win, draw or loss.
- Dashboard
- A Streamlit and Plotly dashboard for team performance. The hosted copy currently asks for a sign-in, so it isn’t linked here.
Learning lab · Education
PyTorch Playground
An interactive PyTorch curriculum in a single HTML file: 21 chapters, checkpoint quizzes and five live labs that run real forward passes and backpropagation in the browser.
Archive
Earlier work
Projects that have been superseded or were smaller in scope, kept here for the record.
-
RIDE-CLI
The first command-line data analysis tool from my master’s project, now succeeded by WRANG.
-
RIDE (Streamlit)
The original no-code platform from my thesis, with AutoML and LLM chat pages. It is preserved in RI2’s Archive folder.
-
American Express default prediction
Gradient-boosted and linear models for credit default prediction on the AmEx Kaggle dataset, built for an advanced data mining course.
-
High-performance sorting visualizer
Parallel sorting algorithms and scaling tests, built for a High Performance Scientific Computing class.