Independent systems builder

Useful systems, carefully made.

I build at the intersection of applied AI, developer tools, and self-hosted infrastructure — aiming for systems the next person can understand, operate, and trust.

See the work GitHub profile

Billy Armstrong smiling, outdoors in front of green foliage

The path so farFocus over time

01

Operations

Running systems other people depend on.

02

Home infrastructure

Self-hosted services, backups, small devices.

03

Developer tooling

Making the inherited system legible.

04

Applied AI

Agents with safer tools and clear failure modes.

05

Building in public

GBrain, Vector Architect, notes along the way.

Three threads I keep returning to: turning information into usable context, giving agents safer tools, and keeping personal infrastructure legible.

01 GBrainKnowledge infrastructureA private knowledge layer for searchable context, citations, and agent workflows that stay close to the source. 02 Vector ArchitectRetrieval architectureA source-led workspace for structured memory, integrity checks, and a clean MCP surface that other tools can extend. 03 Agent operationsRuntime reliabilityPractical work around agent sessions, process lifecycle, and the details that make automation trustworthy. 04 Home infrastructureSelf-hosted systemsSmall, useful infrastructure: network services, backups, and constrained devices that stay observable and recoverable.

Screen recording — GBrain in use (placeholder)

Workbench / home lab photo

Clear interfaces. Honest failure modes. Systems another person can trust.

I like work that sits between the diagram and the lived reality: systems that another person can operate, inspect, and trust long after the handoff.

Focus: AI systems · tooling · operations
Approach: Human-gated · source-led · durable
Now: Building useful things in public

Generative media · a pipeline

Short AI-generated clips (Sora), run end to end through a homemade pipeline: transcribed, word-timed captions rendered and burned in, then compressed to web-ready MP4 and served straight from this Worker — no third-party player.

How it's made

GenerateEight ~10-second square clips from Sora (704×704).
TranscribeMLX Whisper (large-v3-turbo) on Apple Silicon → reviewed word timings.
CaptionWord-timed captions rendered with Pillow, composited via FFmpeg overlay — no libass or drawtext.
Compresslibx264 slow, CRF 24, faststart — 66.5 MB down to 8.7 MB (~87% smaller).
ShipCaptioned MP4s plus VTT/SRT sidecars, checksummed and validated, served as static assets.

“AI did it” — the 3:51 story of building this, narrated over a sped-up agent terminal. The narration voice is my own voice platform: a self-hosted, provider-agnostic speech service (TTS/ASR) with a swappable backend.

The eight clips it produced

Ten-second Sora clips, each transcribed and captioned by the same pipeline — arrow through the set.