Agentic operations platform · 2026
Grivara OPS
A multi-tenant operating system connecting equipment, crews, work orders, maintenance, budgets, and governed agents in one operational model.
TanStack StartTypeScriptORPCPostgreSQL + PostGIS
Available to design and build new agentic products
Marvin.
I build agentic systems that do real work without losing control.
If you're building an AI-native product or a new agentic business, this is the work I do end to end: the data model that becomes the source of truth, the boundaries that make agents safe to trust, and the product people actually use. The case studies below show how, decision by decision.
Discuss your agentic product02
Technical writing
Notes from building with agents.
Long-form writing about agent architecture, durable workflows, evals, and the lessons that only appear after shipping.
13 min
Scoring a funnel-building agent: the judge gets outvoted
FunnelLoops has an agent that turns chat prompts into funnel JSON. Measuring whether it does that correctly took three scoring layers, a dataset that outlives each fix, and an LLM judge that is deliberately the least trusted voice in the room.
14 min
Durable agentic workflows: surviving the crash mid-run
Agents fail constantly. The interesting engineering in Grivara OPS wasn't the prompts. It was making every agent run resumable, budgeted, and verifiable after any failure.
03
Working principles
How I work.
Your problem first, my stack second
Before writing code I get inside the actual operation: where time dies, where money leaks, what a win looks like in numbers. The architecture follows from that, not the other way around.
Working software, early
You see something real running in the first weeks, then we iterate on what actual usage teaches. Roadmaps and decks don't survive contact with the operation. Deployed software does.
No AI theater
If a deterministic rule beats a model, you get the rule. Agents earn autonomy in small, reversible steps behind validation and audit. You're paying for outcomes, not for a demo.
Built to hand over
A feature is done when your team can run it without me: observable, debuggable, documented. I measure success by what keeps working after I leave.