DJ-AI

You may have read a bit about the exciting new project I am working on - Spwig - an all inclusive e-commerce platform.

One thing that has been exploding my ego over the last few months as I progress the development on Spwig is the bewilderment and amazed responses I get when speaking with potential integration partners. They are surprised when they see how much has been developed and how robust and user friendly the interface is - but it's not that which surprises them, it's the fact that the entire platform has been developed by one man (me - yes, I feel alright gloating a little). But in truth, I haven't developed it entirely on my own. I have developed the AI tools TO develop the platform. My latest tool, the help document generation agentic AI system is designed to produce merchant friendly useful guides to ensure there is sufficient, well documented help throughout the admin interface.

This system is a nerds wet dream - First it consists of some pretty cool hardware I managed to score at a decent price on Carousell (a Singapore based second hand marketplace app for those not in the know). It's a server grade desktop running an AMD ThreadRipper 3970X with the biggest heatsink I have ever seen. Coupled with 128GB RAM and two RTX 3090 GPU's all seated on a ASROCK TRX40 Creator motherboard, it's just what I needed for some of my agentic tasks.

The documentation agent I have created, doc-autopilot, is a combination of open-source tools, developed with a DJANGO based admin front end. I've designed it to eliminate most manual writing effort by automatically discovering UI flows, drafting content, translating content and publishing help updates to the Spwig update server so help is always fresh and updated on merchants installations. This cuts down the turn around time for new or updated help documentation dramatically. It's also much better at writing user friendly help than I am… I tend to get too technical when writing, where this system really makes the content readable for anyone.

The orchestration (using celery)  is pretty straight forward albeit with a few checks and balances that I won't dive into here. But in essence it goes a little something like this
 - Discovers the UI components in a Spwig app, reads through js files, models and templates to get an idea of UI flows and what each element does

 - Generates draft help files for each feature (say for example - how to add a shipping provider, or, how to configure your payment gateway etc) - uses Qwen 2.5 7b instruct
 - Goes through the actual admin interface and takes relevant screenshots - uses Playwright automation, MinIO

 - Edits images for fast delivery, blurs PII and adds tool tips to relevant points mentioned from the help files.
 - Combines the draft and images into a rendered template

 - validates (with some human interaction here)

 - Translates (using M2M100)

 - Publishes to the help document repository, fully version aware)

It has been one of the more time consuming tools I have built, but for good reason - I hate help documentation that is useless, out of date and confusing, and I don't want my merchants feeling that way about Spwig's documentation.

In other news, my auditory cortex's encoder / decoder project is still progressing. Training is taking WEEKS!! The decoder is progressing, it's already closing in on 300 epochs (each taking 3 hours so you do the math), but it's a progressive training schedule, so while we reached excellent voice and noise separation for the batch processing phase (roughly Voice SI-SDR value around 13), the streaming phase is much more challenging for the model, which dropped voice SI-SDR below -10, but after several weeks we are back in the positive number, although still a long way to go to get parity with the batch processing SI-SDR performance levels. The noise SI-SDR is a bit disappointing still, increasing but still below SI-SDR -15.. this is inherent in that noise is a lot more varied than voice believe it or not. It'll take a lot longer to get noise to a good standard.