A Kubernetes homelab in a 10-inch mini rack (yes, that's a Hello Kitty flash drive in the photo below), and Vivazo, a Rust game engine I build at night.
Senior or Staff roles in AI infrastructure, platform, or MLOps. Remote-first, US-based.
Frank Cancedda
Software Engineer III at Technergetics and tech lead for our AI infrastructure platform. My specialty is getting modern AI to run in air-gapped, accredited environments, where nothing installs from the internet and every container has to clear accreditation. In practice that means vLLM and Triton serving, Kubernetes, and the dev tooling that keeps all of it livable, for programs in the DoD Platform One ecosystem. On the side I train small models for retrieval and reasoning, mostly to see how far small can go.
Compliance-grade dev environment, one command
Built the Tilt + k3d platform that stands up all 40+ services and jobs of our AI inference stack on a laptop, close enough to production that bugs show up in dev instead of staging. Onboarding used to be half a day of setup docs. Now it's typing `tilt up`.
Modern AI where nothing installs from the internet
Air-gapped networks can't pull from package registries, and every container has to clear accreditation. I packaged vLLM as a self-installing wheel with a guardrail layer, drove NVIDIA Triton through the DoD Iron Bank pipeline into an environment that had no accredited inference server before, and architected the RAG platform around them: PGVector with multiple embedding models and OpenSearch adapters for hybrid retrieval, running fully disconnected.
Release pipelines that replaced hand-run deploy scripts
Rebuilt the release pipeline in GitLab CI: build, test, scan, deploy. Releases went from tribal knowledge in one person's head to something anyone on the team can rerun and audit. I liked the pattern enough that my homelab runs on it now.
I run my own k8s cluster.
That's it in the photo, next to my desk. Flux handles GitOps, Authelia handles SSO, Grafana watches everything, GitLab runs the CI. It hosts real apps, including a fitness tracker with AI meal analysis. Running it on weekends keeps me honest about how this stuff actually behaves when nobody's on call but me.
A 1.6M-parameter policy network that picks a search strategy per query (how to transform it, how to weight retrieval) while staying inside a token budget. Built on my from-scratch PyTorch transformer lab and evaluated on four BEIR benchmarks.
I reimplemented the recursive-reasoning papers (HRM, "Less is More: Recursive Reasoning with Tiny Networks") and pointed them at multi-agent pathfinding, then added per-agent metrics, heatmap traces, and interactive visualizations. Runs against MovingAI benchmarks.
A Rust game engine built browser-first: WASM is a real target, not a port. Has a formal Stable/Beta/Experimental API tier system, so breaking changes are a decision instead of an accident.
- Triton Inference Server
- vLLM · TensorRT-LLM
- PyTorch · from-scratch transformers
- PGVector · OpenSearch · VLM/OCR pipelines
- Kubernetes (EKS, k3d, OpenShift)
- Tilt · Flux · Helm
- GitLab CI/CD
- Authelia / OIDC · Traefik · Istio
- Grafana · Prometheus
- Python — ML, services
- Go — control planes
- Rust — latency-critical paths
- TypeScript / Next.js — UI
- Java — enterprise services