Wen Yang Ho

Engineering leader who still builds.

VP of Engineering at Mortgage Automator, a lending SaaS, in Toronto. On call since 2019. Since late 2025 I also design and ship agent systems myself. The cards on this page are written the way I write runbooks: what triggers it, the steps, the limits.

Download resume (PDF)Email me

LinkedIn GitHub

Activity log

Week of Oct 19 · 160 commits · Webtoon art pipeline 84

1 to 402 commits a weekbefore git historyUpdated Oct 19, 2026

Leading engineering at Mortgage Automator

Sep 2019 to now

Trigger Mortgage Automator is a loan origination and servicing SaaS for private lenders in Canada, the US, Australia and New Zealand. I joined as a senior engineer in 2019 and have been VP of Engineering since February 2024.

Led engineering through technical due diligence for the May 2024 majority recapitalization by BVP Forge, the private-equity arm of Bessemer Venture Partners.
Led engineering through SOC 2: controls, evidence and changes to how we build and ship.
Built hiring, onboarding and a double code-review system while the team grew to about 20 engineers.
Stayed hands-on: rebuilt the Fund engine to at least 5x faster and moved local development to Docker, which cut setup from 1 to 2 days to under 30 minutes.

ScopeUp to ~20 engineers. 375+ lenders and $20B+ funded on the platform as of 2024.

On callGo-to engineer for incidents, disaster recovery and major migrations.

Agent control plane

Running daily · open-source release planned

Trigger About 45 repositories and many Claude Code sessions a day, each one starting from zero and none of them knowing what the others had built.

Session starts: a hook injects a short briefing with the project's status, next steps and what is waiting on me.
Before building anything generic, the agent asks the knowledge map whether another project already has it.
An agent reaching into another project without asking first is blocked once. A retry goes through.
When a turn changed files, the Stop hook asks for a report and the project card updates.

LimitHooks time out fast and fail open. If the server is down, events spool to disk and replay later.

RollbackOne command removes the hooks, MCP server and generated skill and keeps the data.

Multi-agent orchestration

Waverunner is public

Trigger One agent at a time was the bottleneck, and every repo was growing its own runner script.

Declare a wave of tasks once in a config file.
Each task runs in its own git worktree, so agents never step on each other's files.
Mix parallel batches with sequential barriers, and Claude and Codex in the same wave.
Keep the assembled prompts, logs and outputs for review before anything merges.

LimitTasks in one batch must be independent. Anything that depends on another task goes behind a barrier.

RollbackEvery task is a branch in its own worktree. Rejecting one means deleting it.

Cross-family LLM judges

Research

Trigger I wanted to know whether an LLM can rank short-video openings before they are published, and how far to trust it.

Read the evidence first. On the Upworthy headline benchmark, prompted GPT-4 picked the winner 38% of the time against a 33% baseline.
Judge in blind pairs, in both orders, and keep only picks that survive the swap.
Use a panel across Claude, Gemini and Qwen, and count effective votes instead of raw votes.
Test on five real Shorts with known view outcomes before trusting it on my own.

LimitFive videos is a smoke test, and topic familiarity confounds the result. The judge is a filter for wide margins, not a ranker.

DecisionKeep it as a pruning filter; let real retention data decide the close calls.

Agent-safe infrastructure

Public · this site runs on it

Trigger Agents deploy my apps. They should never hold root on a server or the keys to my DNS.

Harden a fresh Ubuntu server in one script: key-only SSH on a random port, firewall, fail2ban, automatic updates and a verify mode.
Run one shared Caddy per server. Projects never open ports; they join a Docker network and Caddy routes to them.
Accept web traffic only from Cloudflare, so the server's real address is never in DNS.
Every script checks state first, can be re-run and takes --dry-run. All of them run as an unprivileged deploy user.

LimitDNS changes run from my workstation with a scoped token. The server never holds it.

RollbackA site-remove script takes a project off the proxy. A full server rebuild was tested in October 2026.