Projects
Open-source software and project write-ups in AI, finance and data analysis.
Forecasting Financial Markets with Timestral
We tested Timestral financial forecasting on a MacBook Air. See how the 28M model performed on stock volume, returns, volatility and local fine-tuning.
Building a tool to compare LLM benchmarks
Compare published LLM benchmarks with Model Atlas. Check the October 2026 dataset, adjust capability weights and compare matching model results.
Perfect 100/100 Lighthouse score with Hugo and plain CSS
My Hugo homepage scored 100 in all four Lighthouse categories on mobile and desktop. I use plain CSS, system fonts and responsive images.
A Buy in the Data, a Sell in the Model
Our Polymarket study checks 24.5 million trade records against settlement and shows how a wrong buy/sell label can distort estimated trading costs.
How I turn my linkblog into a personalized podcast
I turn saved articles into a personal AI podcast with research agents, outside reading and ElevenLabs v4. The post includes the workflow and finished episode.
Choosing a Model on the Pareto Frontier with Jev
A task-level LLM router for pi: Jev classifies the task, a role policy filters the OpenRouter catalogue, and a value function picks from the Pareto frontier.
Degoogling cost me my YouTube feed, so I made my own
I built a self-hosted YouTube subscription feed with public RSS, no API key, Shorts filtering, privacy-enhanced embeds, caching, and timeouts.
The Anatomy of a Decentralized Prediction Market: Notes from the Polymarket Order Book
I study Polymarket microstructure using 30.3 billion order-book events: spreads, depth, wash trading, and a trade-direction error that affects measurement.
F3ED Can't Call an Ace: Fixing a NeurIPS 2024 Tennis Model
F3ED, the NeurIPS 2024 tennis shot detector, mislabels 73% of single-shot serve unforced errors. A 23-line scoreboard OCR reconciler fixes them.
Against All Odds: The Mathematics of 'Provably Fair' Casino Games
A statistical audit of 20,000 crash-game rounds tests the 97% RTP model. At 179 fixed-stake bets per hour, expected loss is 5.37 stakes.
Social Media Success Prediction: BERT Models for Post Titles
Training RoBERTa to predict Hacker News success revealed temporal leakage inflating metrics. How temporal splits, calibration, and regularization fix it.
65% of Hacker News Posts Have Negative Sentiment, and They Outperform
Sentiment analysis of 32,000 Hacker News posts shows 65% skew negative and earn 27% more points. Six transformer and LLM models tested, full data included.
RSS Swipr: Find Blogs Like You Find Your Dates
Build an open-source ML RSS reader with swipe interface. Uses MPNet embeddings and Thompson sampling for personalized feeds that escape the filter bubble.
Building a No-Tracking Newsletter from Markdown to Distribution
Build a privacy-focused newsletter with Python, Cloudflare Workers KV, and Resend API. No tracking and no infrastructure cost. Open source code included.
Visualizing Gradients with PyTorch
Build the right mental model for gradients with this PyTorch visualization tool. 2D surface plots with gradient vectors show the direction of steepest ascent.
Counting Cards with Computer Vision
How I trained a YOLOv11 model to detect playing cards at 99.5% accuracy and built a real-time Monte Carlo blackjack odds calculator using computer vision.
Modeling Glycemic Response with XGBoost
A hands-on project predicting postprandial glucose curves with XGBoost, Gaussian curve fitting, and 27 engineered features from CGM data. Code on GitHub.
Trading on Market Sentiment
GPT-3.5 matched RavenPack's 41% returns in a sentiment analysis trading strategy using 2,072 news headlines. See the full backtest results and comparison.
I Built a CGM Data Reader
Built a CGM data analysis tool with Python to visualize Freestyle Libre 3 glucose data alongside nutrition, workouts, and sleep for endurance cycling.
Crypto Mean Reversion Trading
How I built a crypto mean reversion trading bot using PELT change point detection on Kraken, targeting altcoin price overreactions with automated execution.
My First 'Optimal' Portfolio
How I built Python portfolio optimization tools, tripled the Sharpe ratio from 0.65 to 1.68, and published the results as an academic paper on MPT.
The Tech behind this Site
How this Hugo blog runs on Hetzner with Cloudflare R2 and Workers, self-hosted analytics and newsletters, and Forgejo builds with checks before deployment.