Software and AI consulting · West Michigan · Regulated industries
Two decades in insurance, banking, and retail operations: underwriting, product ownership, transformation programs, and AI product strategy at scale. Now I design, build, and run complete systems directly, for West Michigan businesses and for regulated-industry teams.
Where I'm useful
About me, Eric Gelder
The person who understands the work is rarely the person who builds the system. I'm both.
Requirements pass from operator to analyst to engineer, and intent leaks at every handoff. Arnhem Advies exists to remove that translation layer.
I have spent two decades in insurance, banking, and retail operations: a decade underwriting farm and commercial insurance and writing the methods and procedures an operation runs on, then owning the core underwriting and billing systems of a large insurer through a Guidewire transformation, then leading AI product strategy across insurance and banking in deployments where being wrong had consequences, and serving as product owner on a retail customer-engagement platform. Different industries, identical failure modes: work that lives in people's heads, handoffs that lose detail, and software that almost fits.
Alongside that, for fifteen years I did the invoicing and helped with operations for a small excavating company. That is where the useful education happened: what it costs when a job's details live in one person's memory, what payroll week feels like when the timesheets are on paper, and how much margin leaks quietly between the work getting finished and the invoice going out.
AI has now removed the last barrier between that judgment and shipped software. I design, build, and operate complete systems directly — the work below — with the discipline the regulated years demanded: requirements written down, changes reviewed, correctness proven rather than assumed.
In one line: I've led transformation in regulated companies and kept the books for an operating one, and I can build the fix myself — I'm not a developer who needs your work explained, and not a consultant who needs a team to translate.
The arc
Fifteen years of invoicing and operations for a small excavating company — where the same few people quote the work, do the work, and bill for it.
P&C insurance underwriting, methods & procedures, and a paperless imaging transformation — ten years of living the workflows, not observing them.
Product owner for core underwriting and billing systems through a Guidewire PolicyCenter transformation; built enterprise-wide UAT practice.
Technical director for AI solutions across insurance and banking — GenAI, computer vision, document ingestion — in HIPAA-sensitive, on-premise deployments.
Now I ship complete production systems directly, directing AI systems from several vendors on owned hardware, and I hold every decision. No team, no translation layer.
Services
I start by understanding how the work actually runs. When custom software is the right answer, I design, build, and host it.
I map how the work actually runs, name the real problem, and tell you whether the answer is a process change, a better tool, custom software, or AI — in that order, not the reverse. Where the work is regulated, that includes compliance, auditability, and who is accountable for each decision.
I design, build, and run software for the parts of your business where off-the-shelf tools are too loose, too rigid, or don't talk to each other. You get a working system and someone who keeps it running.
I put AI to work where it saves real time, and I set the rules around it: who decides, who builds, who checks, what it costs, and what has to be true before anything goes live. For regulated work, that means governance you can show an auditor.
The work
Real systems, described plainly. Where something is still in testing, it says so.
Marine forecasting that keeps score against the lake.
113 million historical observations · 29 stations · 2005–2025
A wave and weather site for Holland, Grand Haven, and Muskegon. It reads live buoy data and federal forecasts, then keeps score: every prediction is checked against what the water actually did, so the system can learn where the public forecast runs high or low. Two decades of Lake Michigan history let me test ideas against real rough days. Until the evidence is strong enough, the site says “calibrating,” labels every value as forecast or observation, and never invents a number to fill a gap.
Visit Mitten Waves
Built end to end — the crew's app, the office console, and the database underneath both.
Crews keep working in a basement or a truck with no signal, and the office sees the work when it syncs. Time and job records carry through to payroll accurately. Scheduling, check-ins, approvals, and reporting sit in one place instead of five.
A small product family, shipped and running.
Three live apps for people who run leagues and competitions, on one account system and one design. Proof that I finish, deploy, and run what I build, not just prototype it.
Visit Group CompetitionsThe work spans web apps, business tools backed by real data, internal portals, integrations, practical websites, and operational utilities that touch hardware. The subject matters less than the approach: choose the right design, ship the workflow, and prove it works.
Point of view
AI has made writing code nearly free. What it cannot supply is judgment: knowing what to build, what “correct” means in a domain where wrong has consequences, and when to reach for AI and when not to. A tool that amplifies is only as good as what you give it to amplify — point it at judgment and you get leverage; point it at confusion and you get confusion, faster. Arnhem Advies is built on the judgment side of that line.
Responsible AI Adoption
A different family of AI reviews every change than the one that wrote it, because models from the same family tend to be wrong in the same ways. And an AI's “done” report isn't evidence: I read the change itself. In one trial, four changes arrived with confident write-ups and nothing changed in the code.
A passing check means nothing until I've deliberately broken it and watched it notice. One review found 40 checks reporting success that could not have seen what they claimed to test. The question is always whether the check actually reached the thing it says it verifies.
I plan AI usage as a visible cost, not a surprise on an invoice. Inexpensive models handle routine, bounded work. The capable, expensive ones are reserved for judgment and the parts that have to be right, matched to the risk. You see that plan first.
Building and the full test run happen on hardware I own, before anything goes out. Where data shouldn't leave the building at all, a private AI model runs locally. Fixed, testable rules make the decisions, and the model only explains them.
The process
People decide. AI does the typing. Nothing ships unreviewed or untested. Same seven steps every time, whether it's a website contact form or a system that runs payroll.
How this starts
No retainer, no long proposal, and no commitment to a big build before anyone knows what it should be.
You describe the problem in your own words. I tell you straight whether software is even the answer — often it's a process change or a tool you already own, and that's a fine outcome.
If it's worth going further, I map the workflow, find where the time and margin actually go, and write it down as a plain recommendation. That document is yours to keep — including if you take it to someone else, or decide to do nothing.
Only once we both know what the system needs to do and what finished looks like. Priced against a scope you have already read.
FAQ
Client Intake
Describe the workflow, system, website, or AI adoption problem in your own words — rough is fine.
You'll hear back from me personally, usually within two business days. The first conversation is a conversation, not a pitch — we work out whether there's a real fit.
Your details go to a private intake queue. No mailing list, no drip campaign, never shared.
Prefer email? eric@arnhemadvies.com
