Things I build

My photo library had 100K+ images and I never looked at any of them. Curation needed to feel like a game, not a chore.

A gamified photo library curation app for Apple platforms. Swipe right to keep, swipe left to trash. AI batch cleaning handles the obvious junk so you can focus on the photos that actually matter. Privacy-first — all analysis happens on-device, nothing is ever uploaded.

Built entirely in Swift 6 with zero external dependencies — just Apple frameworks. Runs on iPhone, iPad, Mac, and Apple TV. Features daily and weekly streaks, 12 achievement badges, burst photo cleanup, duplicate detection, and a "Treasure" button for starring the real gems. The AI is deliberately humble — it never recommends trashing photos with faces, unique moments, or existing favorites.

Every project I built kept reinventing the same backend pieces — jobs, tools, knowledge, messaging. Nexus is the shared substrate so the next ten projects don't have to.

A distributed agent platform that unifies ARIA and what was previously Chancery into a single ecosystem. Nexus provides centralized services for tools, jobs, knowledge, communication, and agent runtime — replacing the fragmented backends that grew organically as each project scaled. Seven agents are live, each owning a domain: personal assistant, data pipeline, infrastructure, inference, code maintenance, insights, and social.

Monorepo with workspace packages for core services, API server, MCP server, and a job queue worker with elastic autoscaling across two machines. Config-driven agents run short-lived cycles on schedules or triggers, execute tools, and communicate via inbox-only messaging. Backed by a 191-table PostgreSQL database with 162K photos, 131K knowledge facts, and 58K emails ingested.

Cloud LLM bills were getting silly and I wanted everything personal to run on hardware I own. Local-first, cloud as fallback.

A unified LLM API running on two AMD Strix Halo machines with a combined 192GB RAM and 160GB VRAM, linked by Thunderbolt 5. Forge routes inference across specialized models for text generation, vision, embeddings, transcription, text-to-speech, face detection, and reranking — all through an OpenAI-compatible interface.

Eight always-on models consuming ~65GB VRAM at baseline, with overflow capacity on the second machine for bulk workloads. Powers ARIA, Nexus, and every other project that needs LLM calls. Request logging in SQLite, Prometheus metrics scraped every 15 seconds, and 25 Grafana dashboards for monitoring. The primary LLM provider for everything — cloud APIs are the fallback, not the default.

I want a model that writes like me, not like an assistant. Twenty years of my own messages and emails are the training set.

Fine-tuning a personalized LoRA adapter on Qwen 32B using my own writing — 98,000+ training records extracted from iMessages, Gmail, Google Voice, Instagram DMs, Facebook, and other communication platforms. The goal is a lightweight adapter that captures how I actually write and think.

Extraction and training scripts are complete. The adapter trains in 2–6 hours on local hardware via QLoRA and will plug into Forge as a servable model, giving every downstream system access to a version of the LLM that sounds like me.

Decades of personal audio and video were a black box. I wanted to ask: who said what, when, and to whom — and get a real answer.

A media processing pipeline that ingests audio and video from a VHS archive, Plaud recordings, Google Voice, Facebook, Instagram, iMessage, meetings, YouTube, Apple Photos, and Google Drive. Performs speaker diarization, voice fingerprinting, and face detection to build unified person profiles across decades of personal media.

Every clip gets transcribed, every speaker gets matched to a known voice, every face gets linked to a known person — turning a chaotic archive into a searchable, queryable knowledge base of who said what, when, and to whom.

An excuse to push multi-tenant Next.js routing as far as it would go — six themed brands out of one codebase.

A multi-tenant Next.js app serving six websites for a fictional Christmas elf family from North Pole, Alaska — four character sites, a family landing page, and a management console. One codebase, one deploy, six domains, each with its own theme and personality.

Built as an excuse to explore multi-tenant routing, themed brand systems, and shared component libraries in a single Next.js app.