Products, systems, and experiments. Most of it runs on hardware in my house. All of it exists because I wanted to understand something from the inside.
Everything
MsgStack
Open-source canon infrastructure: one approved source of truth every AI tool can ground on.
An MCP server plus admin UI. Each department owns a canon domain (product specs, positioning, legal, HR policy) with an approval lifecycle, provenance, and alignment scoring. Hybrid knowledge-graph and vector retrieval, and artifact generation that is grounded by construction.
Why I built it
Every team now generates content at AI speed and none of the tools know what is approved and true. I built the layer that answers that question, first inside the day job and then as something anyone can run.
- FastAPI
- MCP
- SQLite
- Knowledge graph
- pgvector
- Open source
GTMBrief
A GTM alignment workspace for product marketers: messaging architecture in, launch assets out.
Draft a messaging architecture, compile the launch assets that hang off it, and keep product, sales, and marketing looking at the same brief. Built as a small paid product to learn what PMMs will pay for.
Why I built it
Most launch chaos comes from a brief that lives in six places. I wanted one object the whole cross-functional team could work from, with a paying customer on the other side of it.
- Next.js
- Docker
- OpenRouter
Syd
A personal AI assistant with long-term memory, running on hardware in my house.
Reachable over Telegram and a web dashboard. Syd remembers, tracks budget and food, reads my calendar and inbox, watches the homelab, and messages me unprompted when something warrants it. Prompts are versioned and evaluated against a regression set before every deploy.
Why I built it
I wanted to know what an assistant looks like when it is yours: your data, your machine, your rules about when it may speak. Most of what I know about running LLM systems in production comes from keeping this one running every day.
- FastAPI
- n8n
- SQLite
- LiteLLM
- Langfuse
- Local + cloud LLMs
- Telegram
The homelab
A repurposed laptop running sixty-odd containers behind Cloudflare Tunnel. This site is one of them.
Payload CMS sites, the newsletter, n8n, Home Assistant, Paperless, observability, backups to a second machine at night. No open ports on the router. A GPU desktop on the same network runs local models.
Why I built it
Because you can just host things. The series on the blog walks through it, from an old laptop to a real domain with HTTPS, and this is what it looks like at the end.
- Docker
- Cloudflare Tunnel
- WireGuard
- Tailscale
- restic
- Uptime Kuma
WOPR
A mobile puzzle-combat game: trace chains on a board while hostile systems telegraph their next move.
Inspired by the 2012 iPhone game WarGames: WOPR. Five-encounter campaign, upgrades, quick battle, a tutorial, and a bot that simulates thousands of games to check the balance before a change ships. Installs as a PWA, works offline, no accounts, no analytics.
Why I built it
I missed the original and nobody was going to bring it back. It was also a good excuse to build something with a spec, a test suite, and no LLM anywhere in the product.
- React 19
- TypeScript
- Vite
- Zustand
- PWA
- Playwright
The content desk
A newsroom of n8n workflows that draft, lint, and stage my LinkedIn writing behind a human gate.
Trend and topic discovery, drafting in my voice from a maintained style guide, a lint pass that quotes the exact span it objects to, then a morning digest where I approve, edit, or kill each piece. The judge earns auto-post only after it agrees with my decisions at a high enough rate.
Why I built it
Automated content is easy. Automated content that sounds like me and never punches down is not. The gate exists because two prompt-only fixes before it drifted back within a week.
- n8n
- Local Gemma
- OpenRouter
- Style lint
- Telegram
Notes from the Machine
The newsletter, plus the pipeline behind it: self-hosted sending, a welcome flow, and a send guard that catches silent failures.
Runs on listmonk with a transactional relay. New subscribers get the latest issue automatically. An hourly guard checks every finished campaign against its recipient count and alerts on any shortfall, because the first real send dropped forty of seventy-three without an error.
Why I built it
I wanted to own the list and the delivery rather than rent them. Owning them meant learning how email fails.
- listmonk
- Resend
- n8n
- Postgres
Second Brain
A personal knowledge graph: entities and relationships extracted from chat history and notes, with vector plus graph retrieval.
Generic entity and relationship extraction over chunked text, local embeddings, one-hop graph expansion at query time. Nightly hygiene merges duplicates and folds pronouns into the right person so the graph stays useful instead of growing noise.
Why I built it
Syd needed memory that is more than transcript search. This is the experiment in whether a graph is worth its upkeep for one person.
- FastAPI
- Postgres
- pgvector
- fastembed
- Gemini Flash
World Monitor
A self-hosted global news and signals dashboard, aggregated and summarized on my own machine.
Real-time feeds, geopolitical monitoring, and AI summaries on one screen. A way to watch the firehose without an algorithm deciding what I see.
Why I built it
Curiosity, and a distrust of feeds optimized for anything other than my attention.
- Docker
- RSS
- LLM summaries
Browser game streaming
Stream games from the desktop at home to any browser, including a work laptop that cannot install anything.
A bridge that speaks to the streaming host on the desktop and re-serves it over WebRTC with a WebSocket fallback on port 443, which survives corporate firewalls and always-on VPNs. Includes a virtual display so a 3:2 tablet gets the right aspect ratio.
Why I built it
The native client was blocked by policy. Working around that turned into an unplanned lesson in display drivers.
- Rust bridge
- WebRTC
- Apollo
- Cloudflare Tunnel
I build the systems, not only the strategy. The point is leverage: AI as infrastructure that changes what a small team can produce, rather than a feature to talk about.