Technical series · AI systems · GPU optimisation · Quantisation
A technical series on serving a compressed reasoning model in vLLM: kernel optimisation, careful measurement, speculative decoding and the final move from CUDA to Triton and Gluon.
10 chapters · 3 companion guides
Read the series
Tutorial · Statistics · Parameter estimation
Connect least squares, measurement uncertainties and Bayesian inference. Change the sample size, polynomial degree and weighting, then inspect the fitted curve and residuals.
Points, noise, polynomial degree and weighting.
Read the note
Tutorial · Sequential estimation · Tracking
Follow an estimate as observations arrive. Tune measurement and process noise, track a moving vehicle, and compare sequential filtering with smoothing and a batch fit.
Observation count and noise assumptions.
Read the note
Tutorial · State models · Nonlinear filtering
Compare position, velocity and acceleration models on a noisy circular path. Explore their assumptions and see how a circular extended Kalman filter models the geometry explicitly.
State model and noise assumptions.
Read the note
Tutorial · Numerical integration · Mechanics
Build orbital simulations from Newton's laws. Switch integrators, adjust the time step and play the orbit while comparing energy conservation with trajectory accuracy.
Integrator, time step and playback.
Read the note
Tutorial · Detector physics · Optimisation
Recover sensor shifts from many tracks with a global chi-squared fit. Explore the effect of track count, measurement error and detector shifts, and learn why reference constraints matter.
Tracks, measurement error and detector shifts.
Read the note