Built CE ('Sea') — a peer-to-peer compute mesh and economy where every node is assumed hostile and the honest majority wins. libp2p networking with NAT traversal, signed capability chains as the only trust primitive, a custom PoW ledger with escrow and payment channels, and Docker/gVisor sandboxing for workload execution. The paid compute economy is live: usage is metered, admitted, and settled on-chain across a multi-machine fleet (Hetzner relay, GPU workstation, laptops, in-browser WASM nodes). Public relay at ce-net.com; packaged for Homebrew, AUR, Scoop, and Chocolatey.
On those primitives: 70+ services rebuilding Google-class infrastructure as mesh apps — S3-class object storage, a Firestore-class realtime document DB, GKE-style container orchestration, pub/sub, CDN, serverless functions, IAM with passwordless device pairing, distributed CI, and an ngrok-style tunnel service.
Designed the OpenJaws cell mesh platform and its underlying Cell Mesh Protocol: TypeScript/Rust distributed runtime where self-discovering 'cells' gossip-sync a shared atlas and expose type-safe capabilities with no bespoke service-mesh configuration.
Serve large language models across machines that were never meant to cooperate. ce-exo distributes models over a heterogeneous fleet by wrapping production engines — vLLM, llama.cpp, exo, Ollama — behind one OpenAI-compatible API, with NAT-traversing mesh transport and capability-gated access: one command deploys an engine to a GPU machine and routes chat traffic to it from anywhere. ce-tabnet takes the same idea to its extreme — a transformer sharded layer-by-layer across browser tabs, pipeline-parallel inference on WebGPU.
LLMs also run inside my production systems: policy analysis over payroll data at a gaming studio (LLM tool-use in Rust), lesson-material generation in Trana, and an AI-policy pre-run scanner for work admitted to the compute mesh. Day to day, the engineering process itself runs on fleets of parallel AI agents with recorded conventions and shared backlogs — 500k+ lines of production Rust and TypeScript shipped solo in the last 12 months.
Proficient in architecting end-to-end solutions across the full stack. Production systems in Rust (Axum/Dominator) for high-performance backends, SvelteKit and React for complex frontends, PostgreSQL for robust data modeling. Integrated Stripe Connect, Slack API, and Fortnox API. Deployed scalable infrastructure using Docker, GitHub Actions, and Cloudflare Tunnels.
Built low-level systems in Rust with focus on performance and memory efficiency: voxel rendering engine (wgpu, ECS, SDF traversal, TAA, bloom) at 60 FPS on mobile SOCs; p95 78ms API response in production microservices; 1.48M messages/sec zero-copy IPC in the Cell runtime.
Shipped Blueberry Browser: an Electron app that exposes an MCP server so AI agents can delegate real-browser tasks — clicking, form filling, reading dashboards — to a live, logged-in Chromium session. No per-site integrations; if a human can do it in a browser, the agent can too.
Self-directed learning focused on understanding technologies from first principles, paired with a habit of shipping. Built and launched Myra / Trana — a multi-platform AI SaaS — solo from a single monorepo; moved Blueberry from concept to daily-use tool in two weeks; open-sourced the Cell Mesh Protocol as a standalone SDK. Committed to clean architecture, maintainable code, and shipping things that actually work.
Mobile-first resale marketplace — roughly 2M lines of Rust. Axum services, React Native client, PostgreSQL, Redis Streams, Stripe Connect, PostNord shipping, and Cloudflare R2. Multi-currency design with a per-listing 'anchor currency' so prices never drift from what the seller chose. Three independent monitoring layers (in-process Sentinel + Cloudflare Worker + Hetzner systemd timer) that fail independently and alert over email and Telegram.
Gaming Studio (Umeå)
March 2025 – Present · Umeå, Västerbotten County, Sweden
Rydenfalk Systems
February 2025 – Present · Umeå, Västerbotten County, Sweden
Self Employed
January 2024 – May 2025 · Umeå, Västerbotten County, Sweden
An AI lesson-material SaaS for Swedish upper-secondary teachers, built solo. One TypeScript monorepo (Turborepo + pnpm) ships four clients from shared packages: web (TanStack Start + React 19 on Cloudflare Workers), landing (Astro on Cloudflare Pages), mobile (Expo, App Store + Play Store), and desktop (Electron). Supabase backend with row-level security, and a full test pyramid — Vitest, Playwright, pgTAP, and Deno — wired into per-app CI/CD.
A peer-to-peer compute mesh and economy: run a node, donate compute, earn credits, spend them on other people's machines. Every node is assumed hostile — Ed25519 identity, signed capability chains as the only trust primitive, a PoW credit ledger with escrow and payment channels, gVisor-sandboxed workloads. 70+ services on top rebuild Google-class infrastructure as mesh apps: object storage, realtime document DB, container orchestration, pub/sub, CDN, serverless functions, IAM, distributed CI. The paid economy is live across a multi-machine fleet; public relay at ce-net.com; installers for Homebrew, AUR, Scoop, and Chocolatey. Grew out of Cell, a zero-copy Rust substrate (1.48M msg/s, ~677ns RTT, embeddable Raft).
Large-model inference across machines that were never meant to cooperate. ce-exo serves LLMs over a heterogeneous fleet by wrapping production engines — vLLM, llama.cpp, exo, Ollama — behind one OpenAI-compatible API, with NAT-traversing mesh transport and capability-gated access: one command deploys an engine to a GPU box and routes chat traffic to it from anywhere. ce-tabnet shards a transformer layer-by-layer across browser tabs — pipeline-parallel WebGPU inference where each tab holds a slice of the model. Also designed a clinical deployment (staff chat on hospital GPUs, PHI never leaves the LAN).
A distributed computing platform where every service is a sovereign, self-healing 'cell'. Cells gossip-sync a shared atlas, auto-discover typed capabilities, and self-replicate to new nodes via Cell.toml manifests. Production cells span AI inference, Docker management, Hetzner cloud provisioning, live trading, and mobile dashboards. Zero service-mesh configuration — add a directory, it joins the mesh.
An Electron browser that AI agents operate like a human. Exposes a Model Context Protocol (MCP) server so any orchestrator (Claude, n8n, cron jobs) can delegate web UI tasks to a real, logged-in Chromium session. No per-site integrations or API tokens — if a human can do it in a browser, Blueberry can too. Used daily for Gmail, WhatsApp Web, and LinkedIn automation.
Rust/wgpu engine built as a Cargo workspace: a reusable engine library (render graph, ECS via hecs, hot-reloaded WGSL shaders, audio, input) drives a hybrid SDF-voxel ray-traversal renderer with TAA, bloom, and atmospheric scattering at 60 FPS on Apple A15 and M1 — plus a separate N-body orbital ('spacetime') simulator running on the same engine.
A browser multiplayer space arena with no game server: every player's browser runs the full authoritative simulation in-tab (Rust → WASM + WebGL2) and peers merge state by quorum — replicated authority instead of client-server. Zero-delay prediction, per-RTT update rates, and population-driven persistence: a lone player's world resets, a crowd's world survives because the crowd is hosting it. The relay is just transport and a warm seed. Deploys are gated on a live browser smoke test of the real production path.
Reference TypeScript/Rust SDK for the Cell Mesh architecture. Provides typed RPC (router/procedure), gossip-based atlas synchronisation, Ed25519 capability proofs, and an ask/tell signal model. Every mesh call is a self-describing Signal forwarded hop-by-hop; cells learn the topology as a side-effect of handling requests. Published as a standalone package — used as the foundation for OpenJaws.