Angelo Zhang

Selected results

• Abuse incident response

  Handled a coordinated fraud campaign against a consumer product from postmortem to prevention: found the root cause in pricing design, shipped the fix, and wrote the review checklist the team still uses.

• Test suites and CI

  Built one production codebase’s test suite from near zero, grew my own platform’s test suite from 223 to over 1,500 files in roughly four months, and made both suites required checks on the merge path.

• Solo operations

  Run the public products on a platform I own as its only engineer, covering releases, incidents, and support.

Experience

1. 2026
   Platform & production-safety engineer (contract)— Consumer AI product

   Contract engagement after a serious abuse incident on a live consumer product.

   • Wrote the incident postmortem and traced the root cause to the product’s pricing design.
   • Shipped the fix and the spending safeguards that closed the exploited path.
   • Set up the cross-project engineering process the team kept after the engagement ended.
2. 2024 — Present
   Founder & principal engineer— LivelyVerse · independent

   Building and operating a multi-product platform: product definition, architecture, implementation, releases, live operation.

   • When an incident lands, the fix ships together with a check that keeps that failure class from coming back.
   • Handle releases, incidents, and support solo.
3. 2024 — Present
   Lead engineer— LeyLine Pro

   Lead engineer on an AI media production platform built around long-running generation work.

   • Closed a class of authorization bugs and added a build check that flags regressions.
   • Lead the engineering behind generation features, billing, and day-to-day operations.
   • Rebuilt the automated test suite from near zero.
4. 2018 — 2023
   Independent product engineer— Various projects

   Independent full-stack product work across client and personal projects.

   • Moved from mobile specialist to full-stack: product, interface, backend, deployment, operations.
5. 2015 — 2018
   Senior Android engineer— Qunar · Ctrip

   Mature consumer travel applications where reliability and release quality mattered.

   • Shipped Android features in cross-functional teams with strict QA, and kept the production-stability habits that came with that.
6. 2013 — 2015
   Android engineer— Sogou

   First engineering job, building consumer Android features.

   • Learned to ship inside an established engineering organisation.

Selected work

• Mendr

  Founder & principal engineer · 2025 — presentWebsite: mendr.dev

  Developer tooling that takes an engineering issue through diagnosis, fix, test, and verification as one loop. Part of the LivelyVerse product family.

  Approach — Diagnosis → fix → test → verify, automated end to end

  • The loop runs end to end without a human relaying between stages.
  • A fix is only offered after it passes the same checks human-written code must pass.
• LivelyVerse

  Founder & principal engineer · 2024 — presentWebsite: livelyverse.com

  Several public AI products, built and operated on one shared platform.

  Approach — Several products · one shared platform

  • Own product direction, architecture, implementation, releases, and incident review.
  • A custom merge queue runs the required checks on the merge path; a separate post-merge audit alerts on defined bypass patterns.
  • Eight cross-environment drift detectors run by change scope and roll up into one required check before merge.
• LeyLine Pro

  Lead engineer · 2024 — presentWebsite: www.leylinepro.ai

  An AI media production platform: one creative workflow across many generation vendors, with jobs that run for minutes.

  Approach — Media generation workflows · multiple model vendors

  • Lead the product’s engineering: generation features, billing, operations.
  • Built its automated test suite from near zero.

Open source

• Forgeline

  Apache-2.0 · 2026 — presentgithub.com/angelozhangai/forgeline

  Governed AI delivery, on rails

  A service that runs AI-assisted delivery as a durable state machine. A requirement moves through review, design, implementation, and verification, and the run survives a restart at any point along the way.

  • Two different model families review each other’s output, so one model’s blind spot does not become the result.
  • Every gate is decided by something deterministic — tests, coverage floors, acceptance contracts — never by a model’s opinion of its own work.
  • It never merges or publishes on its own; those steps wait for an explicit human go.

Capabilities

• Release engineering

  CI design, merge automation, and the checks that decide whether a change ships.

  CI pipelines · merge automation · static analysis · environment consistency

• Model integration

  Serving generation workloads in production: routing, failure handling, cost control.

  Provider routing and failover · cost tracking · long-running job recovery

• Abuse and economic risk

  Free credits, promotions, and pricing can be exploited like any other surface; I threat-model the economics before launch.

  Threat modelling · rate limiting · risk scoring · payment integrity

• Full-stack delivery

  Interface to data model to deployment. TypeScript mostly; Go where the service calls for it.

  TypeScript · React · Next.js · Go · cloud deployment

• Working with agents

  Coding agents do much of the writing; review, merge, release, and production data stay under human control.

  Agent workflows · code review discipline · human-in-the-loop releases

Education

1. 2009 — 2013
   Southwest University of Science and Technology — Bachelor's degree, Communication Engineering

Writing

Occasional notes on release engineering, testing, and running small production systems.

• 2026-08-18

  The gates machine-written code has to pass

  Review doesn’t scale to agent speed. What runs on my merge path instead, and where each check came from.

• 2026-04-20

  Solo, but not alone: working with AI as a teammate, not a tool

  On being the only engineer on a platform of production AI products, and how that became possible.

• 2026-01-15

  What 'production AI' actually costs

  The gap between a working demo and a system you can ship is wider than most plans budget for.