🍔 Burger Logger

A craving list and tasting journal. Write down the burgers you want, then devour them with a 1–5 rating and a note. Every day, 今日汉堡 (today's burger) picks one for you from what it has learned about your taste.

Originally a 2019 bootcamp assignment (Express, MySQL, Handlebars and a hand-made ORM), rebuilt in 2026.

What this project is for

The original taught the basic CRUD loop: form → route → SQL → HTML. This version keeps that loop and uses it to teach three things the other projects in this set don't:

1. HTML over the wire. No React, no Vue, no build step. The server renders pages and htmx swaps them in place.
2. One domain, two doors. People use the web forms; AI agents use MCP. Both go through the same validation and SQL, so they can't disagree.
3. Privacy without accounts. Every browser gets its own private journal from a random key. No sign-up, no passwords.
4. Small ML you can read, with an LLM where it helps. A tiny taste model learns from your ratings; a language model is optional and only processes text.

今日汉堡: today's burger

How the card is made, in three steps you can read in src/taste/:

1. Text → features. Each burger's name and note become tags from a fixed list, with sentiment: "great bacon, a bit salty" → bacon +1, salty −1. A keyword tagger (English and Chinese) runs instantly; if a model is connected it re-reads the text in the background and its validated tags replace the keyword ones.
2. Features → taste. Each rating minus your average is shared among that burger's tags; praise or complaints push a tag one more point. Dividing by count + 2 keeps a tag near zero until it has a few examples (the idea behind ridge regularisation). Weights like smoky +0.8, salty −0.9 are the whole model.
3. Taste → today's pick. Your cravings and 20 classic styles are scored, with a bonus for a patty you haven't had in two weeks. One is drawn with a die seeded by your journal and the date, so it is the same pick all day.

The card is written to feel good, using four small ideas from psychology:

On the card	Example	Why it works
A memory, in your own words	💭 You gave Smoked brisket smash 5/5 "great char". This one is smoky too.	Being remembered feels personal; a generic "you might like" does not.
A sensory preview	Picture it: crispy, lacy edges, cheese melting over the side and a deep smoky char.	Imagining the first bite is part of the pleasure (anticipation, or savouring).
One new thing	✨ Something new for you: your first plant-based pick.	Novelty inside something familiar is exciting rather than risky.
A promise	👌 Easy on the salt: you've said "too salty" twice.	Removing a known worry beats warning about it.

A model, if configured, only rewrites the sensory preview. Until it answers, or if it fails, a template preview stands in, and the card says which wrote it.

See it work

No JavaScript at all	An AI agent logs a burger
Plain <form>s and redirects, so it works everywhere. htmx only makes it smoother.	Real MCP calls against a local server, then the same journal in the browser.

Run it

corepack enable
pnpm install
pnpm dev          # http://localhost:8787, with a local D1 database

Command	What it does
pnpm test	Worker tests inside the real Workers runtime (workerd) with a local D1, once without a model and once with a scripted fake model
pnpm e2e	Playwright: desktop and phone, JavaScript on and off, axe accessibility scans in both themes
pnpm validate	Lint, types, tests, build, e2e, knip
pnpm gifs	Re-records the GIFs above from the live app (needs ffmpeg)

How it works

flowchart LR
  B[Browser<br/>forms + htmx] -->|POST /burgers ...| H
  A[AI agent<br/>MCP client] -->|POST /mcp/key| H
  subgraph Worker
    H[Hono routes] --> V[Zod rules<br/>domain.ts]
    V --> Q[SQL<br/>db.ts]
  end
  Q --> D[(D1<br/>burgers)]

Read the code in this order; it is about 600 lines in total:

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗