Andrew Judd

Full stack developer. TypeScript, Next.js, PostgreSQL.
I build operations software, and the AI pipeline my team ships through.

Farmington, NM Mountain Time Open to contract work

What I do

I build and maintain the operations platform for a title and escrow company. Quotes, invoices, customers, properties, scheduling, dispatch, payments, role based permissions.

Almost none of it is greenfield. I add to a system that people are using that day, so I go one module at a time and check that each step worked before I start the next one. I have been doing this since March 2022, most of it as the only developer on the piece I was holding.

The work I am proudest of is not a feature. It is the pipeline underneath the features.

AI orchestration

I got tired of watching agents rediscover the same traps, so I built the thing that stops it. It is a TypeScript service, 101 modules and about 56,000 lines, that turns a backlog into scoped tasks in dependency order and runs agents against them one at a time.

1,718 work items delivered through it, April to August 2026
101TypeScript modules
56,000Lines in the orchestrator
1,131Dependency edges
158Docs in the context base

Every task starts with the traps already loaded. Each one ships with context pulled from a knowledge base of 158 documents, so an agent picks up a job already knowing that this Oracle package silently swallows its own error, or that this deploy workflow reports success without deploying anything. It is not starting cold every time.

Nothing closes without proof. A task cannot be marked complete on a claim. It needs pasted evidence: a query result, a test run, a real response from the endpoint. Work that cannot be proven stays open. That rule is the whole reason the system is worth anything, because an agent that says it finished is not the same as a thing that works.

I built it because I do not trust unreviewed AI output, and I did not want to spend my attention being the only thing standing between a generated diff and production. So I put the review in the pipeline instead of in my head. The number above is what that buys. The point is not volume, it is that all of it was scoped, reviewed, and proven before it landed.

Built on Claude Code and the Anthropic API. Runs against Oracle, 19 repos, and live production environments.

The platform

19Next.js apps in production
680,000Lines of TypeScript
10Federated subgraphs
408Test suites

Frontend

19 production Next.js apps in an Nx monorepo, about 330,000 lines of TS and TSX. Shared libraries for the design system, auth, the GraphQL client, and an address autocomplete library on Google Places that replaced six different per app versions of the same thing.

Backend

10 federated subgraphs behind an Apollo gateway, about 350,000 lines of TypeScript in NestJS. Money, documents, transactions, people and orgs, property, leads, communications. I hold the line that business logic goes in the backend, not copied into React components and not buried in database packages.

Dispatch and scheduling

29 modules covering scheduling, geographic zone assignment with polygon routing, participants, itineraries, signups, and cron sync. It is field dispatch with a real estate label on it. Routes, time windows, assigned people, and cancellations that have to cascade cleanly.

Money movement

Trust accounting, disbursements, commission calculation, and the US Bank integration behind wires and reconciliation. One bug I found there is a fair sample of the work: a receipt was committing itself before the row linking it to the file got written, so any error in between left an orphaned deposit and a file balance that read too high.

Platform and infrastructure

Moved the platform off VMs onto Azure Container Apps, 25 apps with no downtime, and rebuilt the CI/CD, health checks, and memory alerting to match. Cloud migrations from OCI to AWS to Azure. Ran a database character set conversion. Built a caching layer synced with the database that was faster than our Redis path.

Shipped and running in production

Anthropic API Stripe Twilio Microsoft Graph Google Maps OpenAI US Bank UKG payroll UPS Simplifile FinCEN SendGrid

I also did the Twilio A2P compliance work, and got the marketing system to thousands of sends a second.

How I work

One module at a time, verified before moving on. I do not do big sweeping refactors on a system people are using today.

I will tell you what I think is wrong and why before I change it. If I disagree with an approach I will say so once, clearly, and then build what you decided.

I follow the patterns already in the codebase, including the ones I would not have picked on a blank page. No dead code and no shortcuts. I have inherited plenty of both.

Get in touch

Available for contract and part time work.