Applied AI & IoT Engineer Dylan Capaldi
I build privacy-first AI systems that run entirely on hardware I own — no cloud, no telemetry, no rent. If it phones home, it doesn’t ship.
From the deck plate to the rack.
I’m a U.S. Navy veteran now studying Applied AI & IoT Engineering at FIU. The Navy taught me to trust systems I can inspect, hold, and keep running when the connection drops — and that conviction runs through everything I build.
So I build local-first: the model, the data, the metrics, and the alerts all live on hardware in my own rack. Agents that take real actions, observability you can audit, and inference that never leaves the building.
I run it the way I ran a Navy supply division: phased rollouts, rollback steps written before the change, acceptance tests before a phase closes, and a written record of what was observed — not what was intended.
Hardware you can touch
If I can’t pull the drive and hold it, I don’t fully trust it. Everything runs on owned silicon.
Private by construction
No cloud round-trips, no telemetry, no third-party model keys. Privacy is the architecture, not a setting.
Observable & shippable
Dockerized, CI/CD’d, instrumented with Prometheus & Grafana. If it isn’t monitored, it isn’t done.
Jarvis // multi-zone LLM agent system
Two input paths — Frigate motion events and a Matrix chat bot — converge on one Home Assistant backend and one shared decision log, so you can ask Jarvis why it turned on a light and it reads its own log. Dockerized on Flask + Ollama with iteration caps, JSON-recovery paths, and per-zone timers. Model routing is backed by committed eval harnesses, not intuition.
Tort Agent // local LLM workflow assistant
qwen3.6:35b-a3b on an RTX 4090 via Ollama, with six tools against an Obsidian vault: morning briefs (weather + Canvas deadlines), Apple Notes export, handwritten-note OCR (PyMuPDF + Apple Vision), atomic concept extraction with dedup, weekly course summaries, and vault navigation. Tool-call accuracy measured across models with a committed eval harness.
Jarvis Control Center // full-stack operator console
Real-time console for inspecting Jarvis decisions, overriding Home Assistant, and watching the rack. FastAPI backend (Pydantic v2, 8 live endpoints) serving the decision timeline, Canvas LMS tracking, HA REST overrides, and Prometheus / Node Exporter alerting. Vanilla React frontend with no build step — dark, dense, Linear-style.
Privacy-First Homelab // 7-node AI & IoT infrastructure
7 nodes, zero cloud dependency: RTX 4090 tower for heavy inference and training, TDP-capped RTX 2060 laptop for 24/7 light inference, bare-metal Home Assistant on a ~10W Celeron, UGREEN NAS as Syncthing hub, all behind a Tailscale mesh with multi-gig wired backhaul and UPS-backed critical nodes. Model assignments per node are chosen from eval data, and the engineering decision log is committed alongside the inventory.
MIA-01 Network Segmentation // ops-as-code, in progress
Phased hardening of the homelab network, run like a change-controlled deployment: a repurposed laptop rebuilt on Ubuntu Server as the Omada controller, gateway and managed switch adopted under central management, and a dedicated IoT VLAN whose isolation was verified by test — IoT clients reach the internet and nothing else. Every phase closes only after an executed acceptance test is recorded via PR; rollback steps are written before the change, not after. Repo is private during deployment — happy to walk through it.
vault-sync // local-first cross-device sync
Obsidian vault sync across Mac, Windows, and an always-on UGREEN NAS hub — Syncthing with end-to-end encryption, no cloud subscriptions. A launchd job exports Apple Notes to PDF daily into the vault inbox, where Tort Agent picks them up to OCR.
Languages & Frameworks
- Python
- C (embedded)
- Bash
- SQL
- JavaScript / React
- Flask
- FastAPI
Infrastructure & DevOps
- Docker / Compose
- GitHub Actions
- Tailscale
- Linux (Ubuntu Server)
- systemd
- Prometheus
- Grafana
- Syncthing
AI / LLMs
- Ollama
- Tool-use loops
- Model routing
- Structured outputs
- Eval harnesses
- LoRA / QLoRA
- ONNX
IoT & Security
- Home Assistant
- Frigate NVR
- Zigbee2MQTT / MQTT
- VLAN segmentation
- Omada
- AdGuard Home
- SIEM / IDS
- Zero-trust networking
Logistics Specialist
Supply and inventory operations — 96% inventory accuracy across 200+ line items. Where the phased, rollback-first, record-what-you-observed habit comes from.
Backend, SQL & DevOps with Python · Modern Software Engineering with DevOps
Two bootcamp tracks covering Python backends, SQL, containers, and CI/CD — the foundation the homelab tooling is built on.
Google Cybersecurity Professional Certificate
SIEM, intrusion detection, network security, and Python security automation — the security mindset behind every system I run.
Let’s build something
that runs on its own.
Seeking Summer 2027 internships in applied AI, IoT, and local-first infrastructure — and open to collaborations year-round. Best way to reach me is GitHub or LinkedIn.