AW / CTRL signal: open · summer 2027pittsburgh ET

Andrew Wang. I build applications, delivery systems, and memory tools for AI-assisted engineering.王施元

MS ECE at Carnegie Mellon · AI/ML Systems · Open to summer 2027 internships

665 pull requests · one square eachPXRead
185auto
221ai-review
203look-after
43look-before
13owner
green: merges on evidence aloneamber: needs a human firsthollow: merges on evidence, I read it after

Snapshot · 2026-10-28. Squares show classifications, not all merged PRs. Read the context

›665 PRs classified · snapshot 2026-10-28
about.txt · read-onlyclose ×

About

Hi, I'm Andrew. I'm doing my MS in ECE at Carnegie Mellon, in the AI/ML Systems track. I finish in December 2027. Before that I did a BS in computer science at UC Santa Barbara.

My work centers on three connected projects: PXRead, a reading and library application; Antiochus, the coordination and delivery boundaries around AI-assisted engineering; and Chronicle, a searchable entry point to work history across tools. Together they cover building the application, shipping changes, and retaining what I learn.

This site brings together my projects, internships, and upstream contributions. The PR grid is a dated snapshot, not a live activity feed.

education
MS ECE · Carnegie Mellon
AI/ML Systems · expected Dec 2027
BS CS · UC Santa Barbara
Computer science · graduated 2026
looking for
Summer 2027 internships
AI infrastructure / LLM systemsBackend & distributed systemsPlatform & developer infrastructureForward-deployed engineering
ShenzhenJun–Aug 2026

Longgang Data

AI infrastructure engineering

I deployed LLMs locally and benchmarked them across heterogeneous hardware: H100, AMD Ryzen AI Max+ 395, Ascend 910B. vLLM, llama.cpp, evalscope.

WuhanMay–Jun 2026

BaDaTech

AI application engineering

I replaced a LangChain prototype with a smaller orchestration layer of my own. A typical query went from 2+ minutes to about 20 seconds.

case_files/ · read-onlyclose ×

Case files

My three main projects: a real application, AI delivery governance, and searchable work history.

PXRead

built

A local library and reading system for mixed TXT and EPUB collections, built around durable works, editions, and revisions.

The problem
Source files arrive with inconsistent names, duplicates, uncertain metadata, and competing versions. A flat file list cannot keep reading progress and annotations attached to the right edition when one version supersedes another.
Engineering work
  • A Work/Edition model separates a work from each readable edition; immutable revisions record body changes, and HEAD advancement uses compare-and-set.
  • An import decision flow records source evidence, candidate metadata, and confidence, then chooses append, fork, skip, or review; body files are content-addressed.
  • A single EPUB reading path uses tied annotation positions, with tests for reattachment and visible detachment.
Evidence and boundaries
A working local application with a backend, web and mobile clients, schema migrations, and tests. Source is private; the public discussion covers architecture and implemented behavior.

Antiochus

built

A private control plane for agent-assisted delivery: intake, risk classification, review routing, evidence, and human gates.

The problem
Assisted development can produce changes faster than a person can inspect them. Passing checks and a plausible review do not establish that the evidence applies to the code now under review.
Engineering work
  • Review verdicts bind to an exact commit; a newer commit makes an earlier verdict stale, and routing picks ordinary or senior review from risk.
  • The controller separates task intake and dispatch from authorization, escalates stuck reviews, and surfaces gates that need human attention.
  • Tests exercise controller decisions against fixture repositories and check that parked, human-owned, and out-of-scope work is left alone.
Evidence and boundaries
Working scripts, tests, architecture records, and delivery policies. Several categories are specified but not implemented; this summary does not claim the whole design is deployed.

Chronicle

built

A private command-line entry point for searching, retrieving, and exporting agent history across changing retrieval backends.

The problem
A long-lived archive becomes hard to use when tools and storage change, and raw session output can contain credentials that should not leave the machine.
Engineering work
  • A single standard-library front door maps verbs through an explicit route table to AgentsView while preserving caller arguments and backend exit codes.
  • An outbound redaction wrapper handles text, JSON, and JSONL before output leaves the machine and preserves the backend's exit status.
  • Tests cover routing, quoted arguments, error propagation, legacy-command refusal, installation shims, and synthetic-secret stripping.
Evidence and boundaries
The front door, redaction rules, tests, and backup documentation are present. Memory extraction and conversational privacy annotation remain unfinished; the underlying search engine is third-party.
Historical snapshot of PXRead delivery classifications

This grid describes classifications as of 2026-10-28. The policy and its operation may have changed since.

autoCI is the evidence.185
ai-reviewCI plus a review verdict bound to the commit.221
look-afterMerges on evidence. I read it afterward.203
look-beforeI read it before it merges.43
ownerOnly I can approve it.13

PXRead, Antiochus, and Chronicle are personal projects whose source is currently private. These are selected overviews, not links to private repositories.

journey.log · read-onlyclose ×

Journey

Entries from my own log. The mistakes stay in.

LOG 01incident

The $55 night

An auto-reviewer fired 34 times in under two hours. I switched it to on-demand the same night.

LOG 02rule

A rule caught its own failure

A week after I wrote the verdict-binding rule, it caught a passing verdict bound to the wrong commit (PX-460/461).

LOG 03reversal

Reversed in one day

I put the human review queue in a private repo. A day later I moved it to the work board instead.

LOG 04retraction

I retracted my own design

I proposed some over-engineering and retracted it myself the same day.

LOG 05leak

The leak that built a gate

During a repo split, deleted docs were still readable in two public branches. That's why this site has a privacy-scan gate.

LOG 06trust

Owner sign-off shrank

As the evidence held up, I needed to sign off on less. It isn't a one-way ratchet. The scope can grow back.

14
dead ends in my notes.
I write down what didn't work so I don't try it twice.
upstream.log · read-onlyclose ×

Open source

Fixes I land in the tools I use every day.

47commits landed upstream
8projects
opencodexProvider proxy for Codex CLI and Claude Code18
agentsviewSession search for coding agents10
AzurPilotGame automation toolkit6
Stronghold-ProtocolBrowser tower-defense fan remake5
cockpit-toolsAccount manager for AI coding IDEs3
hstryUnified history for AI agents3
agentmemoryPersistent memory for coding agents1
AzurPilotLauncherLauncher for AzurPilot1

Landed means commits on the project's default branch that I wrote or co-authored. Some maintainers close a PR and carry the change with my authorship kept, so counting merged PRs alone undercounts by more than half. These figures were collected on October 28, 2026; they are not current totals. Commit counts describe contribution activity, not impact or sole authorship.

Selected public contributions

lidge-jun/opencodex #7701 ↗fix(xai): publish the subscription lane's widest context rung for Grok 4.5-4.7

kenn-io/agentsview #2439 ↗fix(parser): read DeepSeek Harness sessions that log job lifecycle events

wess09/AzurPilot #803 ↗fix(config): 移除配置读写的 stdout 副作用

dispatch_slip · read-onlyclose ×

Contact

For internships, engineering work, or a conversation about the tools I build.

dispatch slipAW-2027
toAndrew Wang · 王施元
reSummer 2027 · AI infrastructure / LLM systems
clearancenone
Email is the best way to reach me. There's no phone number here, on purpose.