Jarxi

Prometheus Behind NAT: Scraping Without Open Ports

How a pull-based monitor reaches a machine with no public IP. The tunnel is a phone line, not a mailbox — and nothing exists until someone asks for it.

Prometheus pulls. It sends an HTTP GET to your machine and reads the response. Simple — until your machines sit in private subnets across three clouds, with no public IPs and security groups you’d ......

Posted by Jessie Jia on 2026-08-25

Load Leveling vs Load Balancing: The Fan-Out Pattern

Two queues, one event. Load balancing splits work across identical workers. Load leveling spreads work across time, so a slow consumer never blocks a fast one.

Most queue diagrams look the same: one producer, one queue, several workers. The point is to share work out. More workers, more throughput. Then one day you write a system where the same event goes......

Posted by Jessie Jia on 2026-08-25

Technical System Design Queue Architecture

Speculative Decoding and MTP: Why Guessing Is Free

A forward pass over five tokens costs about the same as over one. That gap is the entire speedup, and MTP is how the model drafts guesses to fill it.

I saw “MTP round-trip” on a checklist for a Megatron conversion pipeline and had no idea what it meant. Two acronyms, one hyphen, apparently important enough that someone had listed it as a thing t......

Posted by Jessie Jia on 2026-08-19

Technical AI Inference MTP GLM

What special: true Actually Changes

It changes nothing when you tokenize. It changes everything when you detokenize, and only if skip_special_tokens is on. Measured against GLM 5.2's tokenizer.

I found the question in my own eval script, which is the embarrassing way to find things. 12text = tokenizer.decode(output[0, input_len:], skip_special_tokens=True).strip()text = re.sub(r"<......

Posted by Jessie Jia on 2026-08-14

Technical AI Tokenizer GLM Reasoning

After the Hexo Skill, a Plugin for Tone

My first Claude Code skill has 700+ downloads on ClawHub, which surprised me enough to write another. This one holds the voices I post in, and it will keep growing.

The first Claude Code skill I wrote was a selfish one. I publish this blog with Hexo, the workflow has enough sharp edges that I kept re-explaining it, so I wrote it down once: front matter templat......

Posted by Jessie Jia on 2026-08-13

Technical AI Claude Code Writing Plugin

What Actually Gets Tokenized in SFT

The model never sees your messages list, only the rendered string. Once that clicks, add_generation_prompt, loss masks and the double-BOS bug all line up behind it.

tok.apply_chat_template(msgs, add_generation_prompt=True) I have copied that line more times than I can count without thinking about what any of it does. The official example writes it that way, th......

Posted by Jessie Jia on 2026-08-12

Technical AI SFT Tokenizer HuggingFace

Multi-level Account and User Design: One User, One Account

A three-level account and user model — legal entity, account, user. Why we rejected this design, then shipped it two months later when one rule changed.

Most billing bugs are not arithmetic. They are identity: whose balance pays for this request. Get the identity model wrong and every table downstream grows a second nullable column and an if. This ......

Posted by Jessie Jia on 2026-08-11

Technical Architecture Database Postgres Billing

Reddit Growth Strategy: Warm-Up, Targeting, and Auto-Fix

How we keep daily Reddit ops on track without tripping moderator wires or API restrictions

Why Reddit still matters for startup storytellingIf you build in public, Reddit is one of the few places where startup founders and marketers actually trade notes in real time. Our current Reddit g......

Posted by Jessie Jia on 2026-03-16

Reddit Startups Marketing Agent Ops