Systems developer · Los Angeles

I build systems that hold up under pressure.

I’m William Nguyen. I design secure, scalable, and robust solutions turning complex infrastructure into software that is dependable, observable, and straightforward to operate.

Explore selected work github.com/William0327

system.profile operational

01Securenarrow interfaces

02Scalabledecoupled workloads

03Robustdurable state

Backend systemsInfrastructureSecurityReliabilityAutomation

01 / Selected work

Engineering, not just screenshots.

Focused case studies on the problem, the architecture, and the trade-offs behind each system.

01 Featured system 2026

Self-hosted AI infrastructure

Audio Station

A lightweight control plane for running GPU audio tools locally without exposing Docker or keeping expensive workloads online.

View source

Problem: Make multiple AI audio tools manageable from one interface while preserving local data and releasing GPU resources when idle.
Architecture: A CPU-only FastAPI controller delegates lifecycle operations through a private, authenticated gateway to isolated GPU workloads.
Stack: Python, FastAPI, SQLite, Docker Compose, NVIDIA runtime, vanilla JavaScript.
Trade-off: Single-user localhost deployment reduces operational complexity, while intentionally avoiding public multi-tenant authentication.
Result: Release candidate with managed provisioning and CPU regression coverage; real GPU acceptance remains an explicit release gate.

02Security subsystemAudio Station

Least-privilege infrastructure

Hardened Docker Gateway

A narrow control boundary that keeps the Docker socket away from the application-facing controller.

Read security model

Decision: Only the gateway sees Docker; the API rejects arbitrary images, commands, mounts, and download URLs.
Controls: Allowlisted workloads, ownership labels, authenticated requests, pinned inputs, and checksum-verified model downloads.
Challenge: Balance convenient web-managed provisioning with a deliberately small, auditable attack surface.
Lesson: Capability boundaries are easier to reason about when dangerous operations live behind a purpose-built interface.

03Developer toolingLocal-first

Visual workflow compiler

Harness Studio

A local-first visual editor for designing, running, and debugging LLM workflows—then compiling them into standalone Python packages with no Harness Studio runtime dependency.

View source

Problem: Make multi-step LLM workflows visual and testable without locking the exported application to a hosted orchestration service.
Architecture: A React graph editor produces a versioned workflow model; FastAPI validates and topologically executes it, streams traces, and compiles the same graph into a deterministic Python package or ZIP.
Stack: TypeScript, React, Next.js/Vinext, React Flow, Zustand, Python 3.12, FastAPI, Pydantic, and HTTPX.
Trade-off: Local-first storage and environment-based secrets keep the tool private and portable, while deliberately targeting trusted developer machines rather than untrusted multi-user hosting.
Validation: Tests cover graph semantics, provider and tool behavior, deterministic output, generated-code parsing, and execution parity between the editor and exported runtime.

02 / Approach

Built around failure modes.

Good systems are not the ones that never fail. They are the ones that fail clearly, recover deliberately, and remain understandable.

Constrain the surface

Small interfaces, explicit permissions, and secure defaults make systems easier to trust.

Make state durable

Important work should survive a restart, expose progress, and be safe to resume or retry.

Design for operators

Useful errors, observability, and predictable recovery paths are product features.

03 / Contact

Have a hard systems problem?

Tell me what you’re building, where it bends, and what reliable looks like. Your message will be stored only in this site’s local SQLite database.

Local storage enabled