Jonathan Armstrong · jarmstrong158
AI/ML engineer, self-taught. Last December I taught myself Python to stop my team counting shipping tickets by hand; that first tool still runs in production. Within months I was shipping a stack of MCP servers on the official registry, a foundation RL model for workforce scheduling, and automation that saves ~20 hours across warehouse logistics. I ship end-to-end and I'm honest about what my tools can't do.
I run operations at a warehouse. Last December I got tired of watching my team count every shipping ticket by hand, twice, once for the pickers and once for the packers, when the numbers were already sitting online. So I taught myself Python to gather them automatically. That became MasterMetrics, and it still runs in production today. From there I kept going, nights and weekends and a lot of it from my phone: reinforcement-learning agents that schedule a real workforce, a stack of Model Context Protocol servers now live on the official MCP registry, and more tools that automate my own job.
The thread through all of it: I understand the domain I'm optimizing, I build the whole thing (model → server → deploy → docs), and I don't oversell — every project's README states what it doesn't do. That combination is what I want to bring to a team.
Run more than one AI coding agent on the same repo and two things break: they forget between sessions, and they overwrite each other. Xylem fixes both — with no server of its own, because git push already enforces "only save if nobody changed this first." A phone and a desktop become equal peers.
Records the why behind decisions in editable JSON, so a fresh session starts oriented. ~92% less context loaded vs. dumping the store; says "I don't know" instead of guessing.
Server-less coordination — claims live on a git branch and git push is the lock. 1,000 simulated concurrent races: zero lost, zero double-granted.
Distills finished work into recallable knowledge and promotes it by earned trust — local → team → org — so the team learns instead of relearning.
I run a warehouse, so I built RL agents that schedule a warehouse workforce — then generalized them. Numbers below are from held-out simulation, labeled as such.
A foundation RL model — variable-shape transformer + LSTM (~18M params), 3-stage curriculum. Pretrain once, fine-tune a new facility in ~30 min, holding ~93% order completion from 5 to 50 workers (sim). The generalization result.
PPO + LSTM with temporal memory: 98.2% order completion across ~9.4 simulated years on a single facility. The validated predecessor that Clark generalizes.
An autonomous Balatro agent: PPO policy + a forward-simulating build planner + tactical heuristics, measured on a held-out fixed-seed harness. Confirmed white-stake wins — an active project, still training today.
The fastest way to prove software matters is to automate your own job. These run in production at my warehouse.
Automates a manual daily shipping-metrics reporting workflow across the team — Selenium + pandas + Excel. In production use today.
Generates Avery barcode PDFs from an Excel pick-list — auto-shrinking SKUs, width-fit Code 39. In production; ships a synthetic catalog so real data stays private.
…and more — full list on GitHub.
Not a bootcamp. A full-time job, a phone, and every spare hour, starting last December. Each month I took on a genuinely harder domain. Tags are color-coded by thread.
A year of GitHub contributions — empty until I started, then relentless.
Monthly contributions — still bending up.
Monthly GitHub contributions.
If you want someone who understands the domain, builds the whole thing, and ships fast — I'd love to talk. Remote, US-based, and available now.