I build AI systems that reason together, repair themselves, and remember.

Systems & experiments

Problem, approach, and current status for each. Longer write-ups live in case studies.

• Quorum

  Multi-LLM consensus reasoning engine

  Problem

  A single language model is a single point of failure, and the dangerous failure is not refusal but confident invention. Betting a workflow on one model also bets it on one vendor.

  Approach

  Convene a council of models on the same question and treat the answers as votes. Cluster them semantically with embeddings so paraphrase counts as agreement, score the level of agreement, and escalate to more expensive tiers only when the cheap ones fail to converge. Where the council genuinely splits, the split is reported rather than resolved.

  Status

  Open source, Apache-2.0. Active.

  github.com/magnetoid/Quorum

• Janus

  Self-developing, self-healing AI agent

  Problem

  Agents fail in ways their authors did not anticipate, and every recovery path written by hand is a path someone has to maintain. An agent that cannot change itself needs a human for every repair; an agent that can change itself can also dismantle the guarantees it was trusted for.

  Approach

  Keep the agent's invariants as an explicit specification separate from its behaviour. When a run fails, localise the fault, propose a change to the agent's own tooling and policy, and test that change against the invariants before it is allowed to take effect. Modifications outside the permitted surface are refused rather than negotiated, and every accepted change is journalled so it can be audited and rolled back. The interesting question is where that permitted surface should end.

  Status

  In development.

  github.com/magnetoid/janus

• torsor-helper

  Local-first MCP server for agent memory and guardrails

  Problem

  A new session is a blank slate, long sessions drift, and agents optimise for what was said recently rather than what is durably true. They rebuild code that already exists and reintroduce patterns that were explicitly rejected, with no audit trail.

  Approach

  An external memory the agent re-reads at session start, ranked by stability rather than recency, fused from vector and full-text search. Architectural decisions are captured as enforceable rules with a drift check, so a rejected pattern fails the build instead of quietly returning. Markdown is the source of truth, which keeps the agent's memory reviewable and diffable like the rest of the repository.

  Status

  Open source, MIT. Local-first — no API key required to run the memory layer.

  github.com/magnetoid/torsor-helper

• Blob

  Agentic-first AI team workspace

  Problem

  Collaboration tools were built for teams of people who occasionally call an AI. That makes the agent a guest: it speaks through a chat box, owns no state, holds no permissions of its own, and forgets the thread between sessions. Every serious piece of agent work then lives in someone's private history instead of in the team's.

  Approach

  Give agents first-class membership — their own identity, scoped permissions, and durable threads of work a human can pick up and hand back. The workspace, not the transcript, is what persists, so a task survives whoever or whatever last touched it. Anything that leaves the workspace stays behind a human gate.

  Status

  In development. Private.

• Morpheus OS

  AI-native commerce operating system

  Problem

  Commerce platforms consume an AI API as an afterthought. Neither a rules engine nor a chatbot is a tool-using agent that can read a store and act on it, and monoliths make extension risky enough that most never try.

  Approach

  A deliberately small core — catalog, cart, checkout, fulfilment — with everything else as a plugin under a modularity contract that tears down cleanly when a plugin is disabled. Agents are treated as a first-class audience: a merchant assistant in the core, an agent kernel with capability scopes, and audience-scoped protocol endpoints so external AI clients can transact without a bespoke integration.

  Status

  Running in production. Private.

• WooPulse

  AI-powered WooCommerce command center

  Problem

  A solo store operator has to be a merchandiser, a copywriter, a pricing analyst, and a competitor researcher at once. The tooling that does this well is priced and scoped for teams of ten.

  Approach

  Connect over the WooCommerce REST API and score opportunities from the first sync. Consensus drafting across several models where judgement matters — copy and outreach — and cheap local models where it does not, so the economics of the AI layer stay sane. Privacy-sensitive stores can route generation through a local model instead of a hosted API.

  Status

  Live. Private.

• Tenso

  Experimental language for neural-network development

  Problem

  General-purpose languages describe neural networks through library calls, so the things that actually break a model — tensor shape, differentiability, device placement — are conventions checked at runtime rather than properties the language knows about. A shape error surfaces as a stack trace deep in a training run instead of as a compile error.

  Approach

  Treat the network itself as a first-class construct rather than a value assembled by library calls, and push shape and differentiability into the type system so that a mis-wired model fails to compile. An exploration of what becomes expressible once that is true, not a bet that it should replace anything.

  Status

  Exploration. Not a product.

Applied work & credentials

Imba Production LLC

Founded 2005, Wyoming-registered. Custom omnichannel CRM and marketing applications, and autonomous agents running lead generation and affiliate systems.

Upwork

Top Rated Plus with a 100% Job Success Score since 2011, across 130+ clients.

Commerce stack

Anthropic/Claude API, Next.js, WooCommerce, Shopify, full-stack e-commerce.

Marketing

Offline through digital — social, SEM, SEO, now AEO. Certified analyst on Google Marketing Platform and Meta. Mostly e-commerce and SaaS.

Cloud Industry LLC

Co-founded 2014, Delaware.