banner VS tesseract-wasm

Turns out I’ve been hacking on https://github.com/rustl3rs/banner; a CI/CD system; for 3 months now. Haven’t been this into an out of work project in decades. I just keep setting really small goals, and if the animals wake me up early, I hack on that goal until it’s ready. Commit; repeat. Very happy with the progress. Some of the current progress has seen the capability to set env vars and use host directories in a task. Next step is to be able to pass the state from one task to another; at which point it will pretty much be able to dogfood itself. 🐶🦴

I'll be continuing to think and hack on Banner; my thoughts, and now code on how a CICD system should work. It has come a surprisingly long way in such a short time. It's certainly not feature complete, but it has been capable of running a test (unit-test; which could conceivably run unit-tests) for a little while. My current focus is constructing event handlers for simple cases, and allowing them to be executed when a matching event is emitted. Thinking to write the handler as a rhai script, and execute the compiled ast. But rhai isn't particularly async friendly, so may have to shift my thinking. Just started looking at Rune and it looks promising.

I’m starting to formulate some ideas I’ve had for a long time around CI/CD. No guarantees it’ll go anywhere, but I find it pretty hypocritical that we advocate for testing the skin off our code before we accept it as production ready, but then basically test in production the pipelines that deliver that automation. There just aren’t easy ways to test a pipeline before you merge it or before you make the change in production. Your pretty much forced into a practice we all dislike. Anyway, thoughts here: https://github.com/rustl3rs/banner

I am working on a new machine learning runtime for CPU inference of ONNX models, focused especially on use in WebAssembly/browsers. It is part of a larger effort to create a modern alternative to https://github.com/robertknight/tesseract-wasm for in-browser OCR. Tesseract's recognition of lines works quite well and uses modern methods (LSTM neural networks), but its detection of text lines, page layout etc. uses heuristics that are brittle.