Consulting
AI implementation for small businesses that want privacy-first, vendor-neutral systems they can own and operate independently.
Forty years in enterprise software (IBM, Disney; IBM Certified Architect) and the past seven years shipping agent systems and on-device AI as an independent developer. I help small businesses turn AI from a scattered set of experiments into working operational systems — assessed, implemented, documented, and handed over so your team can operate them without ongoing consultant dependency.
Small businesses using AI ad-hoc who want to make it systematic. You're already pasting things into ChatGPT or Claude; you want structured workflows that don't depend on remembering to open a chat window.
Businesses with sensitive data. Legal practices, medical offices, financial advisors, service businesses with client data — anyone who can't comfortably send operational information to cloud SaaS. On-device and local-first processing is where I work.
Businesses evaluating multiple AI platforms. OpenAI, Anthropic Claude, Apple Foundation Models, local Ollama, MLX — a vendor-neutral assessment of which fits your business, your budget, and your data policy.
Businesses that want to own their AI systems. Not rent indefinite dependency on another SaaS subscription. If I've done my job, you can operate and extend the system after I leave.
Honest about fit — these engagements are better served elsewhere.
Large enterprises requiring team-scale engagements. I am one person operating a focused practice; projects over roughly $100K or requiring 24/7 managed services belong at a larger firm.
AI marketing content generation at scale. That's commodity work served well by cheaper specialists.
Projects requiring on-site presence more than three to four times per engagement. I'm based in Maine; occasional travel is fine, resident consulting is not.
Written assessment comparing OpenAI, Anthropic Claude, Apple Foundation Models, local Ollama/MLX, and others on capability, reliability, cost, data sovereignty, and operational complexity. Deliverable is a specific recommendation with justification — not a vendor pitch.
Setting up a Mac mini (or similar Apple Silicon hardware) as your dedicated AI operational platform. Local LLM configuration, API integrations, agent workflows, background services. Your data stays on your hardware; your costs stay predictable.
Purpose-specific AI agents for your business functions — email triage, document processing, workflow automation, scheduled analysis. Each agent built on the principle that the agent proposes and you approve. Nothing operational moves until you say so.
Connecting AI workflows to the systems you already use — email, spreadsheets, inventory, CRM, invoicing, document storage. Working with your existing tools rather than requiring platform migration.
Building or reviewing native Apple-platform apps — SwiftUI architecture, Swift 6 concurrency, widgets and Shortcuts, Foundation Models and Core ML integration with fallbacks for devices without Apple Intelligence, and getting through App Store review. The same work behind the apps on this site.
Every engagement includes written documentation, operational runbooks, and hands-on training. The goal is independent operation after handover.
Monthly retainer for businesses that want occasional guidance as needs evolve. Not the default; not recommended unless genuinely useful. If the handover was good, you shouldn't need it.
Scoping conversations, review of your current operations, written assessment identifying good automation candidates, areas requiring human oversight, and a recommended implementation plan with timeline and budget. The deliverable is yours whether or not you proceed to implementation.
Mac mini configuration, agent development, integration with your existing tools, testing with your actual data. Pricing scoped to a fixed budget after assessment. Hourly rate for undefined scope: $200/hour.
Documentation, hands-on training, operational runbook. You and your team should be able to operate and extend the system independently afterward.
Light-touch availability for questions, periodic review, and guidance as needs evolve. Only if genuinely useful.
Each of these is an AI agent I designed and built myself to solve a specific business problem — and in most cases shipped to users. The methodology I bring to client work is the methodology I apply to my own.
Problem: AI features locked to a single vendor’s model.
Swift Package routing model requests across Apple Foundation Models, Ollama, and MLX. 143 passing tests, 14 logged architectural decisions. The vendor-neutral routing pattern.
Problem: Thousands of trail-camera images to sort and label by hand.
Multi-model agent chaining SpeciesNet (specialized vision) into Apple Foundation Models for on-device species identification. Open source.
View on GitHub (opens in a new tab)Problem: Years of disorganized files that nobody has time to sort.
macOS App Store. LLM-driven agent for filename pattern discovery with human-in-the-loop safety: nothing moves until you approve.
Problem: Answering questions from a document collection without sending it to the cloud.
iOS and macOS App Store. On-device RAG with Core ML embeddings and dual-path LLM inference.
Problem: Knowing which AI model actually performs best for a given job.
Python framework published on PyPI. Multi-agent LLM evaluation with validated judge-model methodology.
View on GitHub (opens in a new tab)Problem: Deciding whether local models are good enough to replace paid cloud ones.
Mac app driving larger MLX models beyond Apple Foundation Models’ guardrails. Empirical testing of local-LLM configurations on Apple Silicon.
Problem: A busy inbox that needs triage every day.
Production IMAP email agent running on my own account daily. Transferable to business email triage.
IBM Certified Architect. SmartCloud Orchestrator deployment, Build Forge cloud-migration POC adopted company-wide, DR architecture for 5,000+ users.
First conversation is a no-cost 15-minute call to understand your situation and determine whether an engagement is a mutual fit.
Email: info@robroy.online
I typically respond within one business day. Remote-first; occasional on-site visits for kickoff, mid-point review, and training.