TYRIEN JONES AI ENGINEER · FOREMAN
THE WORK

Portfolio

Everything I’ve built — finished, paused, all dated. Nothing here is dressed up.

LEAD PROJECTOPERATIONAL_CORE
FINISHEDPC Control Hub · PC_HEALTH · ETS2LA
PAUSEDECO · L.I.LY · PC Control Hub
FIRST LINE OF CODEFeb 2026 · 9 months in
SPLITJob · Mine · Wired in
EVERY CLAIM DATED
JOB REGISTER BUILT TO LAST FIRST · PAUSED BUILDS LAST
OPERATIONAL_CORE Business software you buy once and ownMostly original · no AI inside LEAD · v2 PLANNED v1 · 356 TESTS → PC CONTROL HUB Run my PC’s AI from my phoneMostly original FINISHED · PAUSED v1.0 → PC_HEALTH A longevity monitor for Windows PCsMostly wiring good tools together FINISHED · DAILY USE v0.1.0 → ETS2LA Fixing a four-year-old self-driving mod, keeping it in PythonAudit and fix, on someone else’s code COMPLETE 9 FIXES CLOSED → ECO A control plane for governing AI-assisted buildsMostly original, bespoke work PAUSED 0.1.0 → L.I.LY A personal AI stack with a real safety crewBoth PAUSED v0.16 →
01

OPERATIONAL_CORE

LEAD PROJECT v1 · 28 JUN 2026 356 TESTS NO AI INSIDE

Business software a small operator buys once and owns.

The job. Customers, jobs, invoices with GST, payments and backup for a sole trader or small crew. It runs on their own PC, works with the internet off, and there’s no monthly bill.

What’s mine. The design and the rules: no AI inside, every record a plain file on disk, money in whole cents, and one neutral core that wears each trade as a skin.

Where it is. Version 1 works and is tested. Version 2 is planned, with no code yet, until a stability run and a field test with real tradies are done.

Read the full job sheet →
02

PC_HEALTH

FINISHED v0.1.0 IN DAILY USE MIT LICENSED
get_current_healthREAL RESPONSE FROM MY PC · 3 NOV 2026 · EXCERPT
{
  "freshness": { "live": "fresh", "smart": "fresh", "events": "fresh" },
  "data": {
    "smart": { "drives": [
      { "model": "Corsair Force GS", "drive_type": "ata", "smart_status_passed": true,
        "power_on_hours": 18937, "temperature_c": 26,
        "reallocated_sectors": 1, "has_uncorrected_errors": true, "wear_warning": false },
      { "model": "KINGSTON SFYRD2000G", "drive_type": "nvme", "smart_status_passed": true,
        "power_on_hours": 7100, "temperature_c": 41, "percentage_used": 2,
        "available_spare": 100, "media_errors": 0, "wear_warning": false }
    ] },
    "live": { "gpus": [
      { "name": "NVIDIA GeForce RTX 3060", "power_draw_w": 16.96, "power_limit_w": 137.7,
        "throttle_hw_thermal": false, "throttle_hw_power_brake": false }
    ] },
    "events": { "lookback_hours": 24,
      "headline": "Headline: clean window — no critical hardware events" }
  }
}
What an AI gets from the MCP server · drive serials and the GPU’s ID left out · highlighted: the older SSD’s flags

A longevity monitor for Windows PCs.

The job. Watch the health of a Windows PC over time — drive wear, temperatures, GPU, and system errors — and keep a timestamped record so you can see trouble coming before a drive dies on you. Read-only by design: there’s exactly one network call in the whole tool, to a sensor service on my own machine. Nothing leaves the box.

What’s mine. The layer on top of the sensors: polling four sources on a sensible schedule, deciding what counts as healthy versus worrying, writing clean timestamped JSON snapshots — and handing it all to an AI through an MCP server with five typed, read-only tools.

get_current_health get_recent_live get_smart_history get_recent_events list_drives
MATERIALS LISTPYTHON 3.11+
Aggregator + MCP server
Polling, thresholds, snapshots, five tools
MY DESIGN
README + thresholds file
Setup, AI integration, architecture, dependency versions
MY DESIGN
LibreHardwareMonitor
CPU, GPU, board sensors
WIRED IN
smartmontools
Drive SMART data
WIRED IN
nvidia-smi
GPU
WIRED IN
Windows Event Log
System errors
WIRED IN
Model Context Protocol
How the AI reads it
WIRED IN

I didn’t build a hardware monitor. I built the thing that turns four separate hardware tools into one clean stream an AI can read.

github.com/tyrienjones-tech/PC_HEALTH_MCP →
03

PC Control Hub

FINISHED · PAUSED v1.0 · 5 JUL JUN – OCT 2026 AUDITED BEFORE EVERY TAG
PC Control Hub's home screen on a phone: the project, the AI's access level and model, Hub up, and one-tap links to Panel, Scripts, Code, Project shape and Guidebook.
HOME
The Hub's Code screen: the AI says a design audit has to come before any build and that it won't start work on its own, then asks which step to take first. Below, a work ledger lists every command it ran and file it read, each marked done or failed.
CODE · WITH THE WORK LEDGER
The Hub's session menu: new chat, project, save a copy, collect and done, pack up and continue, browse threads, reconnect, stop session, status and guidebook. The footer reads: runs on your PC as you, subscription, no API charge.
SESSION

The Hub on my phone, August 2026

Run my PC’s AI from my phone.

The job. Drive Claude Code on my PC from my phone, anywhere. The phone is a thin web screen; the PC does the work. One live connection with a heartbeat, so the phone shows grey or red within about five seconds if the link drops. The AI on the PC works with full access, the same way I run it at my desk, so all the protection sits at the door.

What’s mine. All of it: ~20,300 lines of Python core and a ~6,700-line phone interface. It has 1,104 commits. 245 are mine; the other 859 were made by AI agents working in roles I set up, and every one is stamped with the role that made it, so the history shows who did what.

Built to recover, with its own system. Before every deploy, the tool takes a backup of exactly the files it’s about to change. If the new version crashes three times in a row within seconds, the watchdog puts the saved copy back by itself, with no AI involved. If it’s down while I’m at the PC, one double-click restores it. Git sits underneath as the history: every change is a commit with the reason written down.

LOCKED DOWN
Network
Only reachable from my own private network, never the open internet
Access
Long random token, checked safely, refuses by default
Secrets
Never in git; every commit is scanned before it lands
MATERIALS LIST
Bridge + watchdog
Python core, rollback, deploy pipeline
MY DESIGN
Phone interface
Web screen, live status
MY DESIGN
Claude Code
The AI doing the work on the PC
WIRED IN
Tailscale
The private network door
WIRED IN

Why it’s paused: keeping my own remote-control layer current was a second job on top of the real work. Claude Code now ships its own remote tools, maintained by the people who build it, so I use those and spend my time on the work itself.

Not finished: the Scripts and Voice tabs are scaffolding, and it’s tested with custom check harnesses rather than a standard test suite yet.

04

ETS2LA — campaign complete

Fixing a four-year-old self-driving mod from the ground up, and keeping it in Python.

ETS2LA is a self-driving mod for Euro Truck Simulator 2. The upstream team is rewriting it in C#, citing performance, maintainability, and code quality. I checked — fifty analysis passes through the codebase — and those reasons don’t hold up. The bugs people actually hit are discipline-and-pattern bugs that come back one-for-one in C# unless you fix the pattern first. This is my own campaign, not a pull-request job: the mod is theirs, GPL v3; the campaign and the argument are mine.

The maintainers won’t merge AI-written code, so the fixes stay in my copy. I finished the campaign anyway, as a learning job done properly from start to end.

“Write like you’re explaining the fix to a foreman, not pitching to a board.”
THE RULE THE REPORT OPENS WITH
BUG LEDGER · 9 CLOSED AND VERIFIED LAST CLOSED JUN 6
ETS-021/022
Arbitrary code execution through the settings fileThe loader ran pickle.loads(eval(value)) on text read straight out of SQLite. Replaced with json.loads; eval and pickle removed from that path. Verified with a hostile-input round-trip test and an in-game window-drag.
ETS-024
A second code-execution hole, in the model-file loadereval() on the contents of a model file. Swapped for ast.literal_eval. Supply-chain reachable — a poisoned model file would have run on the next update check.
ETS-023
Window position that never savedA function that recorded window position was never called. Dead since at least August 2025. Two-line fix, verified in-game.
ETS-005
Map page froze on the second openStale websocket subscription state that never reset on close. Three-line fix, verified in-game.
ETS-025
Silent failures given a voiceThree places in the settings code swallowed errors without a word. One logging import gave each of them a diagnostic line. Verified on the bench and in a two-hour live run.
ETS-026
A retry counter that never countedA database read retried itself with the same counter every time, so the retries were never bounded. Passed the next count instead of the same one. Verified before closing.
ETS-028
A typo’d setting failed silentlyReading a missing or misspelled setting quietly returned nothing, hiding the mistake. It now raises an error right where the typo is. Verified on the bench and in-game.
ETS-030
Lane-change check crashed on an empty routeThe manual lane-change check read the route plan without checking it had anything in it. Guarded against an empty plan; confirmed on a test drive.

How I ran it. Four roles: a Director (Talos) for strategy and scope, an Auto Drive Agent that proposes and lands fixes in the working copy under my checkpoints, an Auto Drive Strategist for a second opinion each cycle — and me as operator, routing between them and driving every fix in-game before it counts as done. Kanban board, append-only JSONL ledger, session logs, handovers.

05

ECO and L.I.LY — paused on purpose

My two biggest builds. They’re where I learned how to structure and run large work, and both are paused: ECO to be rebuilt from scratch with what it taught me, L.I.LY until AI’s laws, rules and pricing settle down. Each has its full job sheet.

PAUSED · 0.1.0 ECO A control plane for governing AI-assisted builds. 279 commits, 678 Rust tests passed. Read the job → PAUSED L.I.LY A personal AI stack with a real safety crew. 159 commits, 609 tests. Read the job →
06

The arc — nine months, dated

I’m showing the early stuff on purpose. These are first attempts, not shipped products. The story is the climb.

FEB 17
Earliest dated file on my machine: an audio archive from before I started coding. (I made music first.)
FEB 19
First Python I wrote: an early audio engine and an LLM bridge script.
FEB 22–28
First real week. Audio tools, a job-hunting web app (the first HuntOS), a couple of unfinished skeletons. Even these shipped with READMEs and troubleshooting docs.
MARCH
Prototypes get serious. Stability Forge gets my first written-down agent crew: Alex, Sarah, and a read-only auditor called TALLAS. That role-separation method has shown up in every project since.
MAR 27
First commit on ECO.
APR 30
Two-role agent rotation set up — Talos and Effy.
MAY 17
Shipped PC_HEALTH. In daily use ever since.
MAY 25
L.I.LY v0.9 tagged after a safety-hardening sweep.
MAY 25–31
ETS2LA campaign: seven fixes closed in the first week, including both code-execution holes.
JUN 6
ETS2LA: last fix closed. Nine verified in total.
JUN 17
First commit on OPERATIONAL_CORE.
JUN 21
ECO paused with Phases 1–3 complete.
JUN 28
OPERATIONAL_CORE v1 shipped: 356 tests.
JUL 5
PC Control Hub v1.0, audited before the tag.
AUG 31
PC Control Hub’s third milestone closes after an independent re-audit.
NOV 3
L.I.LY and the Hub paused. This site moves to Cloudflare Pages.
07

The stack

I’m not claiming to be an expert in eight languages after nine months. I’m claiming I can move between stacks fast enough to ship in them and to audit them.

LANGUAGES I WRITE AND READ
Python
OPERATIONAL_CORE, PC Control Hub, L.I.LY, PC_HEALTH, ETS2LA, most of the March work
Rust
ECO’s backend
TypeScript · Svelte
ECO
JavaScript / HTML / CSS
Web interfaces across L.I.LY, ECO and the Hub, and this site
PowerShell
Backup scripts, the PC_HEALTH event-log piece
SQL
ECO, Founder Brain, and others
C# (read, not written)
Reading a game’s runtime to plan mod work

A lot of what I ship sits on top of open-source software other people wrote. Picking the right tool, pinning a version that works, writing the glue, and documenting how to put it back when it breaks — that’s a real skill.

WHAT I INTEGRATE · NOT MY CODE
AI / LLM
Anthropic’s API, Claude Code, Model Context Protocol
Networking
Tailscale
Backup
Restic
Hardware sensing
LibreHardwareMonitor, smartmontools, nvidia-smi, Windows Event Log
Voice / audio ML
XTTS-v2, faster-whisper, silero-vad
Compute
PyTorch with CUDA
Web frameworks
FastAPI, SvelteKit, Tauri
Storage
SQLite with full-text search, Supabase
Hosting
Cloudflare Pages; self-hosted before that
Quality + security
Playwright, Vitest, ESLint, secret-scanning
THE DISCIPLINE THAT ISN’T CODE

Two backup systems. Daily automated ECO snapshots from April 2026, and a separate encrypted Restic system with a retention policy and a documented disaster-recovery flow.

Verification and decision logs as real documents. The last verified entry is the ground truth. Every decision has a record.

Jobsite paperwork, applied to software. A written brief before the work starts, a sign-off before it ships, and a handover the next person can pick up cold. The ETS2LA campaign above shows the board, the ledger and the handovers.

WORK WITH ME Got a job that looks like one of these?
LET’S TALK ABOUT IT →