Xingji Yan — Software Engineer

I build systems people actually use.

On-call for tenants, retries, and agents that fail closed.

Select a project

20+JASCI · production customer organizations
10s → <1sSearch latency
~95%Fewer DB round trips
CMUM.S. ECE

Now / Looking for

The next system should matter after the demo ends.

I’m looking for software engineering teams building reliable backend platforms, developer infrastructure, or applied AI products where correctness and product judgment share the same roadmap.

How I work: I turn ambiguous workflows into typed contracts, small releases, measurable outcomes, and failure paths the team can operate.
Where I own: I’m most useful when I can follow a problem from the product surface through the service boundary, data model, rollout, and on-call feedback loop.

If your team needs that kind of end-to-end ownership, let’s talk.

Start a conversation →Read résumé ↓

Selected work

CrawlForge

A careers page in. Structured job intelligence out.

JobPosting extractionJSON / CSV20 tests

Live · 20 tests · Updated 2026

Java 21Spring BootSpring JDBCH2jsoup

Problem: Company and ATS careers pages expose the same job data through inconsistent markup and link graphs.
Result: One Careers URL becomes structured title, location, skills, and experience with JSON/CSV export and optional AI matching.
Verified: 20 passing tests · bounded persistent BFS · SSRF-safe redirects · robots-aware retrieval

Open product ↗ GitHub ↗ Case study →

Render may need up to ~60s to wake. View instant proof →

Static proof One Careers URL becomes structured title, location, skills, and experience with JSON/CSV export and optional AI matching.

20 passing tests · bounded persistent BFS · SSRF-safe redirects · robots-aware retrieval

Open full case study →

Catalog Order Service

One last unit. One winning order.

Atomic stockIdempotentRace-tested

Live · 44 tests · Updated 2026

Java 21Spring BootSpring JDBCH2JUnit 5

Problem: Concurrent buyers and client retries can oversell stock or create the same order twice.
Result: With stock = 1, exactly one order wins; a repeated idempotency key replays the original order.
Verified: 44 passing tests · real two-thread race · transactional rollback · bounded webhook retry

Open product ↗ GitHub ↗ Case study →

Render may need up to ~60s to wake. View instant proof →

Static proof With stock = 1, exactly one order wins; a repeated idempotency key replays the original order.

44 passing tests · real two-thread race · transactional rollback · bounded webhook retry

Open full case study →

Clearbay

Financial workflows that stay correct under retries.

IdempotentMulti-tenantRetry-safe

Live · 10 integration tests · Updated 2026

Java 21Spring BootPostgreSQLRedisMCP

Problem: Retries, spoofed tenant context, and concurrent releases can double-bill or leak warehouse data.
Result: Invoice retries replay safely, JWT tenant context wins, and stale concurrent updates return 409.
Verified: 10 passing integration tests · tenant-isolated cache · unique idempotency · optimistic locking

Open product ↗ GitHub ↗ Case study →

Render may need up to ~60s to wake. View instant proof →

Static proof Invoice retries replay safely, JWT tenant context wins, and stale concurrent updates return 409.

10 passing integration tests · tenant-isolated cache · unique idempotency · optimistic locking

Open full case study →

06 more projects

More systems

042026

Grantline

Identity from a signed grant, never a header.

Live · policy-checked grants · Updated 2026

GoReactTypeScripted25519

Live ↗ GitHub ↗ Case study →

Render may need up to ~60s to wake. View instant proof →

052026

Durable Brief

AI workflows that can wait.

Live · durable workflow · Updated 2026

TypeScriptNext.jsVercel Workflows

Live ↗ GitHub ↗ Case study →

062026

PulseQueue

Retries are easy. Correct retries aren't.

Live · lease-safe retries · Updated 2026

TypeScriptReactSSENode.js

Live ↗ GitHub ↗ Case study →

Render may need up to ~60s to wake. View instant proof →

072026

Dockline

The model is never the judge.

Live · 40 eval cases · Updated 2026

PythonFastAPIMCPEvals

Live ↗ GitHub ↗ Case study →

Render may need up to ~60s to wake. View instant proof →

082026

SketchSync

Realtime collaboration over a typed WebSocket protocol.

Live · raw WebSockets · Updated 2026

TypeScriptWebSocketReact

Live ↗ GitHub ↗ Case study →

Render may need up to ~60s to wake. View instant proof →

092026

ResuMatch

Deterministic resume/JD scoring with an optional grounded LLM layer.

Live · deterministic scoring · Updated 2026

ReactFastAPINLPTypeScript

Live ↗ GitHub ↗ Case study →

Experience

JASCI Software Software Engineer 2023 — 2026
A smart WMS used by 20+ 3PL customers and 500+ concurrent users. I took an LLM action layer from a sketch to production — JSON Schema, idempotent writes, audit trails — and I still carried on-call.
- 20+Production customer organizations
- 10s → <1sSearch latency
- ~95%Fewer DB round trips
The same contracts, public: Clearbay
Sonos Software Engineer Intern

Universal Search — Kafka indexing, Redis query cache, search speed +30%.
2022
Tencent Software Engineer Assistant

WeChat identity and payments. Redis −30% DB load. JMeter at 8k users, P99 −30%.
2020

Writing

Collection 01 · Projects show what I built. These notes show how I decide—and link every claim back to code.

View all writing →

019 min read · Updated Sep 6

Crawlers · Reliability

Bounded BFS: How to Crawl Careers Sites Without Crawling Forever

A careers URL is not a job API. The useful system is the one that discovers enough, stops on purpose, and can explain every URL it did or did not fetch. Read essay →

028 min read · Updated Sep 6

Transactions · APIs

Idempotency Is a Product Guarantee, Not an HTTP Header

The header only names a request. The database transaction, uniqueness rule, replay semantics, and side-effect boundary are what prevent a customer from paying twice. Read essay →

038 min read · Updated Sep 6

Agents · Evaluation

The Model Is Never the Judge: Deterministic Agent Evals

Language models are useful actors and useful critics. They are the wrong final authority for tenant isolation, tool schemas, write permissions, and other invariants that must not drift. Read essay →

048 min read · Updated Sep 6

AI workflows · Durable execution

Human Approval in Durable AI Workflows

A human-in-the-loop button is not a safety boundary if the process forgets the wait, the approval has no identity, or the model can call publish by another route. Read essay →

Failure modes

Happy paths are easy.

I design the other paths.

Four production guarantees, made visible: concurrency stays correct, authority expires, retries stop, and model output is never trusted by default.

01Race Two buyers, one unit Catalog Order Service Both target stock = 11 CREATED / 1 CONFLICT The database admits exactly one winner; final stock is zero, never negative. Inspect the case study → 02Expire Grant past TTL Grantline TTL reaches 00:00EXPIRED / DENIED A signed grant is still rejected when its five-minute authority has ended. Inspect the case study → 03Fail Bounded retries PulseQueue 1s / 2s / attempt 3DEAD LETTER The job stops after its retry budget and becomes operator-visible work. Inspect the case study → 04Hallucinate Untrusted tool call Dockline SCHEMA / AUTH / EXECUTESCHEMA REJECTED Deterministic schema and authorization gates decide whether execution is possible. Inspect the case study →

What I care about

I ship the unglamorous systems that keep operations alive — then I put an agent on top of them, with guardrails, so people can talk to the warehouse instead of clicking through twelve screens.

01 Correctness

Retries should not double-write. Concurrent checkout should not oversell.

Seen in Catalog Order Service
02 State

Processes should survive the request that started them.

Seen in Durable Brief
03 Failure

Failures should be visible, bounded, and recoverable.

Seen in PulseQueue
04 Humans

Automation should know where its authority ends.

Seen in Dockline

Java/Spring · React/TypeScript · AI systems · M.S. ECE · Carnegie Mellon