---
title: "十方空烬 · Ember ☄️ (@widechaos) | Hashnode"
description: "I build AI, backend and cloud systems, and share practical work on agents, voice, real-time interaction, open source and automation."
url: "https://hashnode.com/@widechaos"
---

@widechaos

# 十方空烬 · Ember ☄️

@widechaos · Joined October 2026

I build AI, backend and cloud systems, and share practical work on agents, voice, real-time interaction, open source and automation.

### [A 267k-parameter intent classifier trained from scratch, running in the browser](https://widechaos.hashnode.dev/a-267k-parameter-intent-classifier-trained-from-scratch-running-in-the-browser?utm_source=hashnode&utm_medium=feed)

Imagine two support messages. One says the cash machine kept my card; the other says the card came back but I got no money. Both mention a cash machine and a card, and both describe different problems

1h ago · 5 min read [widechaos.hashnode.dev](https://widechaos.hashnode.dev/a-267k-parameter-intent-classifier-trained-from-scratch-running-in-the-browser?utm_source=hashnode&utm_medium=feed)

### [Reviewing accepted items does not reveal what a filter missed](https://widechaos.hashnode.dev/reviewing-accepted-items-does-not-reveal-what-a-filter-missed?utm_source=hashnode&utm_medium=feed)

Updated 2026-11-02: version 0.2.0 adds logical precision and recall bounds on the finite input set. The original 0.1.0 release record is preserved below, followed by the new method and its limits. 0.1

1h ago · 4 min read [widechaos.hashnode.dev](https://widechaos.hashnode.dev/reviewing-accepted-items-does-not-reveal-what-a-filter-missed?utm_source=hashnode&utm_medium=feed)

### [Separating a Hit from Support: What an Offline Evidence Finder Actually Guarantees](https://widechaos.hashnode.dev/separating-a-hit-from-support-what-an-offline-evidence-finder-actually-guarantees?utm_source=hashnode&utm_medium=feed)

Two questions get run together. Did this sentence appear in my documents? And does this document support the claim? The first is decidable by a string slice, the second is not. Offline Evidence Finder

1h ago · 4 min read [widechaos.hashnode.dev](https://widechaos.hashnode.dev/separating-a-hit-from-support-what-an-offline-evidence-finder-actually-guarantees?utm_source=hashnode&utm_medium=feed)

### [Splitting by Declared Groups: A Conditional Alternative to Row Mixing](https://widechaos.hashnode.dev/splitting-by-declared-groups-a-conditional-alternative-to-row-mixing?utm_source=hashnode&utm_medium=feed)

Dividing a collection of records into train, validation and test is often done by shuffling rows and cutting twice. Shuffling by itself is not an independence test and does not make rows independent;

1h ago · 7 min read [widechaos.hashnode.dev](https://widechaos.hashnode.dev/splitting-by-declared-groups-a-conditional-alternative-to-row-mixing?utm_source=hashnode&utm_medium=feed)

### [My LoRA improved routing and made JSON less reliable](https://widechaos.hashnode.dev/my-lora-improved-routing-and-made-json-less-reliable?utm_source=hashnode&utm_medium=feed)

By Renchao Wu My small routing model became much better at returning the expected fields after fine-tuning. It also became slightly worse at returning valid JSON. On 576 held-out synthetic requests, s

Oct 15 · 7 min read [widechaos.hashnode.dev](https://widechaos.hashnode.dev/my-lora-improved-routing-and-made-json-less-reliable?utm_source=hashnode&utm_medium=feed)