Joseph Diaz
Senior Workplace Technology Lead: AI & ML. Building things during the day, running a homelab and an autonomous ops agent after hours.
A technology enthusiast driven by where things are headed. I build and ship AI and ML systems as part of Major League Baseball Technology, and outside of work I run the same kind of experiments at home — most recently a homelab that mostly runs itself.
- based
- New York, NY
- focus
- AI/ML enablement, workplace technology, platform engineering
- stack
- Python Datadog Machine Learning Claude Code Prompt Engineering
- education
- B.S. Computer Information Systems, Lehman College
- Engineered an end-to-end observability pipeline in Datadog, aggregating metrics from distributed hardware assets into real-time dashboards that cut response times to critical events.
- Used Python and data analysis to turn raw ticketing logs into recurring-incident insight reports for cross-functional engineering teams.
- Applied foundational ML (linear/logistic regression, clustering) to historical operations data, exploring predictive models for system downtime.
- Maintained technical documentation and data schemas across the incident management ecosystem to keep automated reporting accurate.
- Parsed, cleaned, and analyzed historical account-provisioning logs with Python to optimize access management workflows.
- Used AI platforms (ChatGPT, Claude, Gemini) to accelerate script development, draft technical documentation, and troubleshoot network connectivity logs.
- Managed asset and inventory datasets, tracking hardware/software lifecycles through structured database management.
- Conducted social media vetting and background analysis for candidates under strict confidentiality standards.
- Supported database management and cross-functional coordination for onboarding review.
FashionVibes is an AI wardrobe app: photograph your closet, AI tags every item (category, color, style, season), and it builds full outfits entirely from what you already own for any "vibe" you ask for — Y2K, streetwear, business casual — each with a written rationale and a matching Spotify track that autoplays. Free/Pro subscription via Apple In-App Purchase.
Background removal on every closet photo was the app's single largest per-image cost, run through a paid third-party API. I built a self-hosted replacement: an open-source segmentation model accelerated on a home Mac Mini's GPU, served persistently and exposed securely over a Cloudflare Tunnel — no ports opened, no exposed home IP. The backend tries it first and falls back automatically to the paid API on any failure, so reliability doesn't change even though the primary cost does.
Auth is email/password with rotating refresh tokens; both AI steps (clothing tagging and outfit reasoning) run through a swappable provider interface, so the model behind either one can change without touching the rest of the pipeline. Source: github.com/origine33/fashionvibes-cc (private repo).
| service | state | up |
|---|---|---|
| media-stack | up | 4d |
| sync-service | up | 4d |
| homelab-agent | up | 39d |
The stack runs on a small home server instead of a dedicated NAS OS — Apple Silicon rules out the usual x86 bare-metal options, so everything lives in Docker on macOS instead. Config and secrets stay on the internal disk; bulk media sits on external APFS drives.
Since July 2026 it's been looked after by an autonomous agent built on Claude: it runs from a set of runbooks (health check, diagnose, restart, backup), reads and writes its own memory of what it's learned about the system, and fixes routine problems without being asked — it only pauses for anything destructive.
Built on top of an existing PrinterLogic install for a printer fleet at work — no new hardware, no new paid software, just automation layered on data the fleet was already producing.
-
01
Found the local monitoring sourceLocated the PrinterLogic desktop client already running and its SNMP log, which tracks whether each printer on the network is reachable.
-
02
Polled for live detailWent beyond up/down status by polling each printer directly for toner level per color, paper tray level and size, and error codes.
-
03
Built a Datadog dashboardShipped the metrics to Datadog: total healthy printers, toner by printer/color, paper by tray, and a breakdown of error types across the fleet.
-
04
Added an automatic alertA Datadog monitor flags any printer the moment a cartridge hits empty or drops below a low threshold.
-
05
Automated a daily reportA scheduled script checks every printer each morning and posts a PDF report to Slack listing what needs toner or paper.
-
06
Built a Slack bot for on-demand answersIT staff can message it directly — "toner", "status", a printer's name — and get a live answer instead of opening a console.
-
07
Automated weekly maintenanceThe fleet power-cycles automatically every Sunday morning, with a Slack summary confirming how many printers rebooted successfully.
-
08
Documented everythingA reference guide covers every script, schedule, and setting, so the system can be maintained or extended without starting from scratch.
None of it required new infrastructure — it reused data the fleet was already producing. The same pattern (parse what already exists, automate the repetitive checks, add a simple way to ask) applies to almost any recurring IT task: server health, license renewals, backup checks, ticket triage.
Open to new roles in AI/ML and platform engineering.
- runescape
- voldimordt →
- basketball
- guard · 3-point shooter
- pokémon
- water type · squirtle