Computer Vision, understood.
Then shipped.
Condados teaches computer vision the practical way: interactive demos, the real math, and runnable Python & C++ for every core idea. Every benchmark here names the hardware it ran on.
Read the latest blog posts, or follow along by email.
Rotate, Scale, Shear: Images as 2×2 Matrices
Lesson 1 of Image Alignment. A 2×2 matrix is where it sends the two axes; its determinant is how much it scales area and its SVD is rotate, stretch, rotate. The one thing it cannot do is make parallel lines meet. On a pickleball court, the sidelines are 20.35° apart in the photograph, so the best 2×2 misses every corner by 60 px.
Browse by topic
Everything on the site, from how images work to models running on edge hardware.
Benchmarks & edge deployment
Measured on real hardware, with the code to reproduce every number.
Edge Deployment Marigold V2 at four bits: what quantizing a 20-billion-parameter depth model costs
Marigold V2 ships as a 17 GB inference job, but the released checkpoint is a rank-128 LoRA over a frozen 20.4 B backbone, so the backbone is swappable. Four configurations on one rented A40: going 4-bit costs 10% of the time and saves 65% of the memory, picking GGUF over bitsandbytes costs another 1.87x and saves nothing, and streaming the weights from host RAM costs 3.33x.
Edge Deployment Build a computer-vision plugin for OBS Studio, and measure what every frame costs
A step-by-step OBS Studio video filter in C++ on OpenVINO, split so the model swaps without touching the plumbing, and measured stage by stage: getting pixels off the GPU, where inference has to live so OBS never drops a frame, and what the trip from camera to virtual camera costs.
Models & Detection Reproducing a published pupil-diameter model, and running it in the browser
An independent run of Shah et al.'s released PupilSense checkpoints over all 424,000 EyeDentify crops, then the whole thing converted to ONNX and put in a browser tab: what the quantisation costs, what the preprocessing port costs, where the compute goes, and what I would change to make it both cheaper and more trustworthy.
Learn the fundamentals
Explainers and interactive demos, with the real math and runnable code.
Fundamentals Affine vs Perspective Transform
Lesson 2 of Image Alignment. Written on (x, y, 1), a 3×3 matrix can translate (affine, six numbers) and, once its last row is allowed to vary, make parallel lines meet (projective, eight numbers). On a pickleball court, an affine map fixed by three corners misses the fourth by 3.50 m; a homography through four lands the held-out centre-line corner 0.47 cm from the rulebook.
Fundamentals How to Compute a Homography from Four Points
Lesson 3 of Image Alignment. Each point pair gives two linear equations in the nine entries of H, and the answer is the singular vector of the smallest singular value. Normalising the coordinates drops the condition number from 53,466 to 6.4, and on this court changes the answer by under 0.2 cm. What decides the error is where the pixels are: one pixel is 0.6 cm of court at the near baseline and 8.1 cm at the far one.
Fundamentals Homogeneous Coordinates: Why Vision Adds a 1
Lesson 1 of Camera Geometry, and the prerequisite of Image Alignment. Add a third coordinate and a point becomes a ray, a line becomes a 3-vector, and both the line through two points and the point where two lines cross are one cross product. Measured on a pickleball court: a corner 18.5 px outside the frame, found anyway, and two sidelines that are parallel on the ground meeting 2,252 px to the right.
Fundamentals Bird's-Eye View and Panoramas: Warping with a Homography
Lesson 5 of Image Alignment. Pushing each pixel forward through H leaves holes in 56% of the far court's bird's-eye view; pulling each output pixel back fills every one exactly once, and matches OpenCV's warpPerspective to one grey level. Interpolation decides what the pulled-back value is. Then a median of 31 warped frames removes the players, who cover 7% of the court in one frame.
Evaluation How many good parts does it take to beat a System One AI model? Between one and sixteen, on VisA
System One models, the idea TypeSafe launched with Jev in September 2026, answer a question with a decision in a single pass. Jev-Omni, an open one, and the untouched Gemma 4 12B it was built from inspect all 2,162 VisA test images without seeing a good part. PatchCore, the standard industrial anomaly detector, is given k good parts per product: at 256 pixels it passes Jev-Omni at k = 8 and Gemma 4 at k = 16, at 512 pixels at k = 1 and k = 4. Also: the base model beats its System One fine-tune, and a day-0 demo on a lime line.
Fundamentals What Is a Homography?
The unit overview: what a 2×2 matrix can do to a picture, why a 3×3 is the smallest map that fits a plane seen in perspective, how four correspondences determine it, how to survive the wrong ones, and how to resample the image through it. Measured on one frame of a pickleball final: an affine map misses the court's far sideline by 3.97 m where a homography misses it by 6 cm.
Get new posts by email
New explainers, interactive demos, and benchmarked projects as they publish, plus short notes on computer vision and the AI world.