Uses — juliardi

The tinkerer page. What I sit at, what I run, and what I reach for. Items marked [fill in] are placeholders until I list the real hardware.

Desk

Editor: VS Code for most days, with an AI coding assistant always open beside it.
Terminal and shell: zsh, tmux for long-running sessions on remote boxes.
Notes: Obsidian. This site’s content is written in the same vault it is published from.
Diagrams: Mermaid in Markdown, because a diagram I cannot diff is a diagram I will not update.

Inference box: a single-GPU machine for running quantised models offline. It exists so I can demo air-gapped systems without asking anyone for a data centre.
Kubernetes at home: a small cluster for rehearsing the Helm and Kustomize layouts I use for clients.
The CCTV lineage: my first homelab project was a CCTV built from a Go binary, Pion WebRTC and a webcam. Everything since has been the same instinct with better hardware.

Claude: my default for reasoning through architecture and for long agentic sessions on a codebase.
Cursor: for in-editor edits when I already know what I want.
Codex: for batch changes I can review in bulk.
How I use them: as production tools, not toys. Every workflow that runs unattended has evaluation and a way to stop it.

Languages: Go for services, Python for anything that touches a model.
Data: PostgreSQL first (pgvector when vectors are small enough), Redis for the hot path, Qdrant when retrieval is the product.
Serving: vLLM on GPUs, Ollama for laptops and demos, KServe and Kubeflow when a pipeline needs to outlive a project.
Cloud AI: Google Vertex AI (Gemini), Amazon Bedrock, Azure OpenAI / AI Foundry, behind model-agnostic gateways so a client can bring their own model.
Observability for LLMs: Langfuse and Arize Phoenix for traces and evaluation.
Voice: LiveKit Agents over WebRTC, Whisper for STT, a TTS pipeline that can run on-premise.
Infra: Kubernetes with Helm, Terraform for the cloud pieces, GitHub Actions for CI.