Over the past several months I've designed, built, and shipped everything below — from real e-commerce stores taking live payments, to a private AI research platform and media library, to the server infrastructure it all runs on. One rented dedicated server, a home GPU rig, and a lot of shipped code.
Nearly everything here is self-hosted: I run the servers, the databases, the reverse proxies, the mail, the backups, and the deployments myself.
Before Zeus existed, I built the platform it stands on: a fully self-hosted AI research environment where every query, document, and conversation stays on my own servers. It's the kind of stack most people rent from Google, OpenAI, and Notion — except here every piece is open-source, wired together by hand, and gated behind my own single-sign-on.
This stack is what makes the rest of the portfolio possible: the book translator borrows its vector memory, Zeus grew up on its retrieval engines, and every project gets model routing and observability for free. It runs 24/7 on my primary VM and has for months.
Zeus ingests almost anything — YouTube videos, podcasts, PDFs, EPUBs, web articles, voice memos, even live streams — transcribes audio on a local GPU, and indexes every chunk into both a semantic vector store and a keyword search engine. You can then ask the whole library questions in plain English and get answers with citations that jump to the exact audio timestamp or book paragraph.
The architecture is split across two machines: the server VM runs the API, search, and studio, while a home GPU rig does the heavy lifting — WhisperX transcription, PDF-to-markdown conversion, audio stem separation, OCR — coordinating over a job-lease heartbeat protocol so either side can fail without losing work. Built and verified through eleven development phases.
SRM is a real-time platform built on WebRTC for hanging out in rooms: voice chat, text chat, and — the part that makes it special — a co-presence engine that keeps video playback in sync for everyone in the room. It integrates directly with Zeus, so a film from the media library can be watched together, with everyone's player locked to the same frame, while talking over it in voice.
It runs in production today across two VMs on my infrastructure — including a dedicated private backend VM reachable only through the reverse-proxy front door.
The foundation is a single dedicated server running the Proxmox hypervisor, divided into seven VMs, each with one job: a primary application box (the AI research stack and Zeus), a hardened commerce-and-mail box that runs the live stores, a thin reverse-proxy edge, dedicated site and real-time-media VMs, and an isolated box for the collaborator project. Memory is deliberately overcommitted and balloon-tuned across the fleet, with capacity decisions — which VM gets RAM, what gets powered down to free it — made and documented like a small ops team would.
Public traffic enters through one thin edge VM running Caddy, which terminates TLS with automated Let's Encrypt certificates and reverse-proxies to backend VMs that live on a private bridge with no public IP at all. The edge also NATs the backends' outbound traffic, so a new site or service costs zero new IP addresses: add a DNS record, add a proxy block, done. Sensitive workloads that need their own identity — the mail-carrying commerce box, the isolated collaborator box — keep dedicated public IPs instead, each bound through the host's virtual-MAC routing.
I run my own mail server — DKIM-signed, deliverability protected by giving it its own dedicated IP away from experimental workloads — handling real customer order mail for the stores. Identity across the AI stack is centralized in self-hosted Authentik single-sign-on, with services gated at the proxy layer so nothing private is reachable without auth. Management access rides a Tailscale mesh VPN rather than open SSH ports, exposed services are minimized, and the boxes that need hardening get fail2ban and unattended security upgrades.
The server fleet handles serving; the heavy AI compute happens at home on a GPU workstation. It runs the Zeus worker — WhisperX transcription, PDF-to-markdown conversion, audio stem separation, OCR — pulling jobs from the server over a lease-and-heartbeat protocol, so the rig can go offline without losing work and the server never depends on it being up. The same rig hosts local LLM inference (Ollama and LM Studio serving open-weight models), a hardened sandboxed model container, the model-swap proxy, and the AI video-generation experiments. The split keeps monthly server costs flat: rent cheap CPUs for uptime, own the GPU for the expensive work.
Every VM backs up nightly to a Proxmox Backup Server running on a repurposed laptop at home, reached over the VPN mesh — an off-site copy of the entire fleet, on hardware I physically control, with the job configured to auto-include any new VM. When things break, they get real incident response: the morning the backup server fell into a crash-loop, the root cause (a permissions bug on a temporary filesystem) was diagnosed, fixed, and written up as a post-mortem the same day. The known weaknesses — single disk, no second off-site copy — are documented too, because pretending otherwise is how data gets lost.
A full e-commerce storefront selling framed, canvas, and unframed prints (plus mugs and cards) from ~90 original photographs. Customers get a per-photo crop tool, size options filtered by print quality, and Stripe checkout; orders flow automatically to a print-on-demand partner with a cash-flow gate so nothing goes to production before funds settle. Margin tracking, discount codes, transactional email, and an admin console — all built from scratch.
A multimedia archive for a 1960s–2000s local band: history and timeline, member bios, digitized cassettes and reels with in-page audio players, reunion video, podcast interviews, and a moderated fan memory wall that accepts photos and voice memos. A built-in merch shop sells shirts and mugs through Stripe with print-on-demand fulfillment. Backend services are small Python daemons I wrote — contact, cookieless analytics, memory moderation, and order handling.
A pipeline for researching local businesses with no (or bad) web presence, building them a fast, clean demo site, and publishing it at its own subdomain as an opener for outreach. Includes a photo-optimization toolchain, a researched prospect database of ~200 verified local contacts, and repeatable one-shot site builds.
I wrote a native Android camera app that streams live video from repurposed phones into a self-hosted NVR with AI motion detection. The first-generation app runs on a Moto G5 Plus watching the garden right now; a second-generation port for a Pixel 4a includes a fully spec'd Power-over-Ethernet build for permanent outdoor mounting.
The public monetization surface for the Zeus studio pipeline: a static gallery of captioned, watermarked GIFs with a lightbox viewer and copy-ready direct links that play inline on Discord and Telegram. First-party, cookieless analytics (self-hosted SQLite + GeoIP) — no third-party trackers.
A self-contained Next.js application collecting a set of smaller AI builds in one place, running as its own service on my infrastructure.
An experiment in trustworthy AI coding: a deterministic plan → implement → verify loop where the language models have no authority over correctness. Tests are authored blind, validated against planted pass/fail fixtures before any live run, and a cross-family judge model is only allowed to rule on what the test attestations already prove. Three complete projects were built end-to-end through the loop as proof, driven from a web interface I built for it.
Securities fraud leaves a paper trail in EDGAR, but it hides in the filings nobody reads. Forensic Harvester reads them at scale: a zero-cost deterministic screen sweeps thousands of companies (2,300+ in one sweep) using only free SEC data — no API keys, no AI spend — and only the leads that survive triage ever touch a language model. A three-axis scoring engine separates procedural manipulation from accounting strain from ordinary distress, so honest strugglers don't get ranked like frauds.
My role was advisory: system architecture, the evidence-spine design (no claim survives without backing evidence), and infrastructure and deployment guidance. I wrote none of the application code — the build credit belongs to my collaborator, and it runs in production on an isolated VM in my fleet.
Splits a French book into segments, translates them through a large language model, and keeps a vector-database "translation memory" so terminology stays consistent across chapters. Fully resumable, parallelized from an 8-hour sequential run down to about an hour, and finished chapters upload themselves to my cloud with shareable links. Currently working through a real book.
A local LLM proxy running as a system service that lets my coding agent swap between DeepSeek, Kimi, Mistral, or a fully local GPU model with a single click from a desktop GUI I built — including a card that drops everything back to the default provider. Alongside it: a curated meta-prompt library with a production evaluation pipeline, and a hardened sandboxed container for running open-weight models locally on my GPU.
Ported Anthropic's open-source legal skill suite — 12 practice areas from commercial contracts to litigation to AI governance — onto a different self-hosted agent framework, rebuilding the skill-routing layer so all 150 skills auto-discover and route correctly, with role gating and per-team practice profiles.
A working wrapper around the BioCLIP vision model that identifies plants and animals from photos, with both a command-line mode and a drag-and-drop web interface, running on my own GPU.
An old LG V30 running a terminal environment with a share-sheet hook: tap "share" on almost any video link on the phone and it downloads locally — including sources protected by bot checks, handled via TLS impersonation and token solving. Entirely offline-first once media is saved.
Testing how far an 8 GB consumer GPU can be pushed for AI video work: running diffusion video models locally to restyle footage into a 1970s comic-book look, benchmarking VRAM use and generation times, and documenting what works at this hardware tier.
A 2D isometric battle engine with initiative-queue turn order, elevation-aware pathfinding, data-driven abilities, and enemy AI — architected so the entire game core runs headless under automated tests, with rendering kept as a thin layer on top.