I don't stop at “look what AI can do.”
I want to know what it can reliably do for the business. Then I build the workflow, the application, and the way people will actually use it: agents, MCP connectors, and automations that moved real work.
Curiosity is broad. The standard is specific.
I work across agent systems, model evaluation, connected business software, creative tools, and interface design. The question is not which technology is most fashionable. It's which combination solves the problem without creating three more.
I keep research and production apart. A local test is not a rollout. A historical simulation is not a prospective prediction. A model that produces a plausible answer is not automatically a trustworthy system.
The marketing brain stays in the room: audience, positioning, conversion, distribution, and the economics of the work.
Read how I think about evidenceAI adoption is people work, too.
At a large commercial manufacturer I built the marketing team's first multi-agent framework on Claude and MCP, wired agents into the content library, the events platform and the CRM, and shipped internal apps that replaced inboxes and spreadsheets. Sales reps across the company used the tools, and the team leaned on far fewer outside agencies and software subscriptions.
The harder part was people. I coached marketers and executives from one-off AI requests to building their own automations. The real bottleneck in most organizations isn't model quality. It's nobody trusting an agent enough to hand it a real task, and clear scope plus a visible review step fixes more of that than any model upgrade.
Read the in-house case studiesWhat I actually reach for, and why.
Frontier models
Claude
Primary model for agents, MCP servers, and this site's own chat
OpenAI / GPT
Comparative benchmarking, some client-side integrations
Gemini
Runs the nano-banana image pipeline behind this site's own art
Vertex AI
Google's model + agent hosting for enterprise deployments
OpenRouter
Model routing and fallback chains, including this site's own chat
AI-assisted development
Claude Code
Directs full builds end to end, this site included
Cursor
AI-native editor for faster iteration on generated code
GitHub Copilot
Inline completion across day-to-day work
Automation & agents
n8n
Self-hosted workflow automation
Make
Visual automation for lighter-weight client workflows
Zapier
Fast integrations where speed beats customization
AI creative
Nano Banana 2
Gemini's image model, generated every hero image on this site
Midjourney
Concept and mood exploration for brand-heavy work
How agents stop copy-pasting and start working.
The Model Context Protocol is how an agent gets a controlled way into the systems a business already runs on. These are the kinds of servers I've built for real teams, described without the vendors or the internals.
A content-library server
Built for a sales-enablement content platform: finding and managing assets, content operations, batch jobs and reporting, plus a command-line toolkit and app templates.
An events-platform server
Generated from the platform's API spec and secured with OAuth2, with operation indexing, token caching and the smallest scopes that would do the job.
CRM data connectors
Let an agent read lead records, clean and dedupe them, and write them back, so routing works from data it can trust.
- Defined scope
- Start with a bounded task and make permissions explicit.
- Reviewable workflows
- Keep consequential actions behind a human review step.
- Evidence before output
- Ground answers in approved sources and make uncertainty visible.
How an agent goes from idea to everyday tool.
- 01Pick a job worth automatingStart from where the team loses time or misses revenue, not from what the model can do. If the job isn't worth a person's attention, it isn't worth an agent's either.
- 02Map the systems and the permissionsWhich platforms hold the data, what the account can really do, and who has to sign off. Documentation is where verification starts, not where it ends.
- 03Build narrow toolsMCP servers shaped around tasks: find, inspect, prepare a change, confirm. Five tools that finish the work beat fifty that expose every endpoint.
- 04Write the contractScope, capped outputs, a review step for anything that leaves the building, and a hard stop. The model works inside the contract, not around it.
- 05Test against the real thingRun it on the actual account, data and edge cases. A passing demo is a small conclusion. Failures get written down where the next decision gets made.
- 06Roll it out with peopleTrain the team on what the agent may touch and where they sign off. Trust is the real bottleneck, so the rollout is part of the build.
I run my own work on agents, too.
Nova is my coordinator agent. It has a defined role and personality, hands work to other agents, and leans on self-hosted n8n for the workflows that should run the same way every time.
Underneath it I'm prototyping Fleet Kernel, a control layer with durable tasks, scoped tools, model-routing rules and budgets, where a task only counts as done when there is evidence it got done. Novabot puts persistent bots and group conversations behind a chat and a companion iPhone app.
Client systems get the same rules I hold my own to: a scope, a review step and a hard stop.
Read the full case studyBuilt. Building. Investigating.
Reviewed
A snapshot of my working project list. Status reflects the project record, not a live uptime monitor. Private systems stay private; the useful story does not have to. Every app-like project also lives in the Software console.
Applied research / product developmentSaturviewPublic beta
A football intelligence product combining sourced data, rankings, season simulations, and an assistant constrained to supported records.
Current boundary: Retrospective research, real outcomes, and prospective picks remain separate. No validated sportsbook edge or predictive-accuracy guarantee.
Explore the projectDesign research / agent briefsForm AtlasLive tool
A source directory and brief builder for turning design references into actionable direction for agents.
Current boundary: A reference and briefing product, not a promise of automatic faithful recreation of any site.
Explore the projectCreative AI / multimodal developmentDaydraftPrivate working app
Guided Agent and direct Studio modes for image, video, audio, and text workflows, with a persistent local library and remix history.
Current boundary: Production deployment and billing remain pending in the project record. Not presented as an available public service.
Agent research / reliable executionFleet KernelPrototype
An event-driven control layer with durable tasks, scoped tools, model-routing policy, budgets, and evidence-gated completion.
Current boundary: Prototype tests do not establish universal cross-harness production readiness. Existing production agents remain separate.
Interface research / capture and editingAgent CanvasPrototype
A browser-first capture and editing workspace with an inspector, variations, undo and redo, save and reopen, and safe static HTML export.
Current boundary: Verified against supported static-layout fixtures. No public deployment, provider integration, or arbitrary-site fidelity claim.
Agent tooling / evaluationAstra SuperharnessResearch
An installed Codex-plugin experiment exploring structured agent execution and reusable working methods.
Current boundary: An HTML benchmark did not complete successfully through the plugin. No efficiency, savings, or superiority claim is attached to it.
Marketing intelligence / workflow designContent strategy agentIn development
Separates information gathering from strategy generation, with dated inputs, idea expiry, and correction-driven knowledge updates.
Current boundary: Some source connections remain blocked. The design records missing evidence rather than treating it as complete coverage.
Agent interfaces / orchestrationNovabotIn development
Persistent bots, bounded workspaces, group conversations, streaming chat, and a companion iPhone interface.
Current boundary: An active build, not a public production service or verified universal orchestration layer.
Creative systems / model integrationOpenVisionPrivate deployment
An owner-operated creative studio with model discovery, a generation library, and remix workflows across media types.
Current boundary: Owner access only. Public demos do not expose private media or access credentials.
Business software / conversational editingSmarterSites StudioIn development
A website-editing portal pairing conversational changes with a live preview and review-gated publishing.
Current boundary: Planned import and messaging features are not represented as shipped integrations.
Also on the bench
Public writing lives on the AnovaGrowth blogOpen research lab
ActiveA research publication where every post is written by a real agent session, run through a multi-agent pipeline of builder, scientist, reasoner, and inventor roles, with a full audit trail.
Design systems
PublishedNERVE v3, a light-first design system for AI products, and NovaAI, a React and Tailwind kit with 467 icons and design tokens reverse-engineered from OpenAI's app design language.