Joseph Landry

I do the driving. AI does the typing.

I'm Joseph Landry, a systems administrator. I write the spec, AI writes most of the code, and I test it until it holds up on real data.

Most of what I make starts one of two ways - a process I got tired of doing by hand, or something I care about that's clunky or abandoned and deserves better.

01How I think

The rules I build by

These come from fixing the same problems more than once. They show up in everything I make, from a document generator to a home firewall.

  • Simple beats clever.

    I'll take a little less automation over a lot more architecture.

  • Low variance.

    A process should work in almost any environment without per-case tweaking.

  • Search before you build.

    Reuse what exists. Write something new only when nothing fits.

  • Nudge, don't block.

    When the data is messy, the tool asks you to fix it instead of refusing to run.

  • Same input, same output.

    A generated document shouldn't depend on who ran the tool or which AI built it.

  • Read-only by default.

    Tools look before they touch. Nothing changes until a person says yes.

  • Accuracy where mistakes are expensive.

    A bad config mapping costs hours of troubleshooting later, so that's where I spend the extra checks.

  • Built for the next maintainer.

    Documented and structured so another person, or another AI, can pick it up cold.

02How I build

Spec first, then let the machine type

I'm strongest at logic and at seeing a problem from an odd angle. AI fills the hands-on coding gap. The loop looks like this.

  1. Find the friction

    Something takes too long, breaks the same way twice, or nobody has bothered to fix.

  2. Write the spec

    What it does, what it must never do, and how I'll know it worked.

  3. AI builds, I review

    Agents write the code. I direct, read the diffs and approve.

  4. Prove it on real data

    A quick go or no-go test before I sink more time in.

  5. Ship small, keep patching

    Other people use it, report what breaks, and I version it forward.

03What I build

AI tooling for real work

At my day job I build tools for other technicians and the people they support. The specifics belong to my employer, so this is the shape of the work rather than the details.

Skills that turn a process into a tool

In daily use

Packaged AI skills that draft scopes of work, engineering plans, bills of materials and client status reports. Each one produces the same structure no matter who runs it, and the status reports can run on a schedule as drafts for review.

Connectors that let AI read business systems

In daily use

MCP connectors that give AI read access to the ticketing and documentation systems a tech lives in.

  • Read-only by default, with per-user keys. Write tools exist but stay off unless someone turns them on.
  • Built by CI and published as releases, so installs update themselves.
  • Sensitive fields are excluded on purpose. The connector reports that a secret exists, never what it is.

A fleet of single-purpose agents

Running

A set of AI bots where each one has one job - tracking my time, catching messages that slip through, scouting a project before it starts, watching our own tools for known vulnerabilities. An orchestrator bot helps me build and maintain the rest.

  • A desired-state baseline rebuilds a bot's machine after a wipe, with minimum versions pinned for security.
  • Proposed fixes wait on an approval list until I sign off.
  • Teammates don't need to know which bot does what. One front-door bot routes the request.

A config migration mapper

In production

Takes one firewall vendor's config export and produces a mapping workbook and a click-path build guide for another. The output is checked three ways - against the target's API schema, against an independent re-read of the export, and with a traffic simulation. Other techs use it on live migrations and I patch it from their reports.

04Projects

Things I build on my own time

These are mine, start to finish. They're the ones where I ask "can this be done?" and then go find out. See them all in the portfolio.

Debugging the tool that helps AI read my code

Fixed upstream

My game's AI workflow runs on vexp, a context engine that gives coding agents a pre-indexed, ranked view of a codebase so they don't have to grep around and burn tokens. On a 120,000-line C# project it wasn't returning the right context, so I took it apart and worked with its developer directly. Over two days and four releases I sent four reports, each with a repro package and debug traces.

  • Search found no C# code at all. I queried the index database directly to prove the data was there and scoring well, then traced the failure to how the search layer handled C# names like CombatActor. The developer adopted the fallback I suggested - promote strong keyword hits when the semantic score comes back empty - and shipped fixes in the next release.
  • The fix worked for some classes and not others. I ran a passing class and two failing ones through the pipeline side by side and showed exactly which step dropped them.
  • 22,000 call-graph edges were collected but never read. A changelog said this was fixed. The debug trace showed the impact query never touched that table, because the editor and the indexer wrote function names in two different formats.
  • A view everyone would read backwards. I documented that the blast-radius section listed what a class calls, not what calls it, so nobody would trust it for the wrong job.

"This is one of the most useful bug reports we've received... seriously, thank you for the depth here; it made our job easy."

Nicola Alessi, vexp developer

A firewall that was holding back

Running

I rebuilt my home network on a Sophos XGS 116 running OPNsense instead of its stock firmware. Along the way I found its SFP port is 10G-capable hardware that the vendor software limits to 1G. Repurposed consumer routers and spare enterprise switches fill out the rest.

Game preservation

In progress

Rebuilding games so they keep running on modern systems and work offline. One of them is a game my wife loves but rarely has time for. An offline build means she can pick it up and put it down with no pressure to keep up, and we can still play together over the home network.

MX64 to OpenWrt flash guide

Written

A lean how-to so someone else can repeat the conversion without me in the room. Only the steps that matter.

05Activity

What the last year looked like

Every square is a day, and the brighter it is, the more I committed. Most of it is the game, which lives in a private repository, so the work shows up here even though the code doesn't.

Commits in the last year
-
Busiest day
-
Lines of C# in the game
121,060
C# files
1,023
System specs
20
Test files
200+

Game numbers as of November 2026.

06Background

Where the experience comes from

I run client engagements from scoping through discovery, documentation and deployment.

Identity and endpoint

Entra ID, Intune, Windows 365 and Active Directory. Passwordless rollouts, Cloud PC provisioning, and getting legacy apps to install where they were never meant to.

Networking

Meraki builds and site cutovers, VPN hub designs, and firewall migrations from Sophos, Palo Alto and Fortinet.

Files and migration

SharePoint migrations in both directions, tenant splits, and NAS setups with delta sync and snapshot-based recovery.

Writing it down

Plain-language scopes of work, phase-by-phase work plans, and handoff docs written so the help desk never has to call me.

07Contact

Say hello

If you're working on something that needs a systems person who can also get AI to do the heavy lifting, I'd like to hear about it.

I'm just as happy to talk about the things I build for the fun of it - lately, that's mostly a world that runs on its own.