Skip to content
View liu-x27's full-sized avatar
2 followers · 2 following
  • Arlington, VA

Block or report liu-x27

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
liu-x27/README.md

Xinyu Liu

M.S. Software Engineering Systems at Northeastern University, Arlington VA (Dec 2028). B.Eng. Software Engineering, Xi'an Jiaotong University. Looking for a Summer 2027 SWE internship — DC metro or remote. A Fall 2027 co-op works too.

Most of what I build ends up being about the same thing: systems that fail without raising. A dictionary that silently outranks real words with typos. A decoder that loops forever instead of erroring. A labelling tool that saves a failed model call as a label. A safety gate that goes on reporting success after the component it depends on has quietly stopped answering.

Each project below has a page that tells its story in a few minutes, and a README that says what was measured, on what, and what was not tested.

mini-claude-code

Doesn't ask about wc -l. Does ask about rm -rf. A readable coding-agent harness on the Claude API, whose risk gate asks a local model four narrow questions about each shell command. On 153 commands it had never seen: 0 of 76 unsafe ones cleared, 38% of the safe ones run without a prompt.

Project page · Code · TypeScript · MCP · ACP

XavierJev

Ask for Y or N. Read the ratio, not the prose. The decision layer behind that gate, as its own library: yes/no, one-of-n and rubric questions read off one token's probabilities. Of the 1,181 real commands it cleared, every one read by hand, 6 should have been asked.

Project page · Code · TypeScript · Claude Code plugin

Lexica

recieve still resolves. It just ranks below receive. A 3.4-million-entry offline English–Chinese dictionary and a lecture captioner I use daily, on Windows and Android from one source tree. One rule ranks down the 3.24 million entries no source vouches for; captions run at 9.1× real time.

Project page · Code · Electron · SQLite · whisper.cpp · Kotlin

Crowd Annotation

Label with a model. Keep what it said separate. A text annotation platform where model drafts are stored apart from human labels. Rewritten after migrating v1's own database showed that only 91 of the 3,063 labels its README called reviewed one at a time were made at a human pace.

Project page · Code · TypeScript · Postgres · React

spire-jev

A narrow win, not a win rate. A bot that plays Slay the Spire 2 in the real game: all five characters, whole runs, no human input. A simulator of the game's combat, checked against the game card by card, searches each turn in well under a millisecond at the median. It has won twice at Ascension 10, and 0 of 90 on the seeds no change was tuned on; its pages say both. Players can run it inside their own game from the Steam Workshop (Jev 自动爬塔).

Project page · Code · TypeScript · C# · Steam Workshop

TypeScript · JavaScript · Python · React · Node · Electron · Kotlin · PostgreSQL · SQLite

Reach me at liu.x27@northeastern.edu.

Pinned Loading

  1. XavierJev XavierJev Public

    A decision layer for agent control flow — yes/no, choice and rubric questions answered from one token's probabilities, measured against labelled sets — with a Claude Code permission hook and a trai…

    TypeScript 2

  2. lexica lexica Public

    Offline dictionary for Windows and Android, plus offline lecture captioning on Windows. 3.4M entries, zero network.

    JavaScript 2

  3. mini-claude-code mini-claude-code Public

    A readable coding-agent harness on the Claude API: agent loop, tools, Claude Code-style permission rules and hooks, MCP, skills, subagents, compaction, ACP. An optional model-scored risk gate (Xavi…

    TypeScript

  4. spire-jev spire-jev Public

    A bot that plays Slay the Spire 2 in the real game, all five characters, whole runs: a combat simulator and search for each turn, rules around the fights. Two ascension-10 wins so far (Ironclad see…

    TypeScript

  5. crowd-annotation-platform crowd-annotation-platform Public

    Assign, annotate, review, and export text datasets, with model drafts stored separately from human labels.

    TypeScript

  6. kydlikebtc/awesome-jev kydlikebtc/awesome-jev Public

    1207 public resources for Jev, TypeSafe AI's System One decision model, indexed by decision pattern. Source citations, dated link checks and scheduled call-site text checks; runtime and performance…

    Python 596 20