Ionel Roiban

Agentic AI · Data science · Palm Beach, FL

Agents that earn their keep

Twenty years of production engineering, now pointed at agent fleets, MCP servers, and calibrated data systems. I design them, build them, and run them — with evals and human gates, not vibes.

Get started

{Agents} {MCP} {Pipelines} {Evals}
How I work
$claude --fleet ship "wire my CRM to an agent that files the paperwork" ⧉ Copy

Services

Agents up front, engineering underneath

Six practices that compose: agentic AI and MCP on top, data science and evals in the middle, twenty years of production infrastructure at the bottom.

Agentic AI & MCP

Agent fleets with shared memory, MCP servers that expose your systems as tools, and human-approval gates on anything with side-effects. Agents that survive contact with production.

You get — a working agent wired to your real systems, with an off switch.

Explore agentic AI →

LLM Systems

RAG over private knowledge bases, local-first deployment on Ollama and llama.cpp, model selection with evals, and context engineering that keeps answers grounded.

You get — a model that knows your business and can prove it.

Explore LLM systems →

Data Science & Analytics

Calibrated forecasting, behavior analytics, BI dashboards in Looker, Tableau, and Grafana, automated alerting, and metrics design — on a Cisco-certified data stack.

You get — numbers your board can lean on.

Explore data science →

AI Workflow Automation

n8n pipelines with LLM steps, self-hosted marketing automation (Mautic plus AI), and process monitoring — the ops glue that keeps automations from rotting between runs.

You get — workflows that run themselves and report in.

Explore automation →

Cloud & MLOps Architecture

Multi-region AWS with 99.99% documented uptime, HIPAA- and SOC II-shaped designs, blue-green deploys, and model-serving infrastructure — twenty years of web platforms underneath.

You get — infrastructure an auditor and an agent can both trust.

Explore cloud & MLOps →

Prediction Markets & Quant

Kalshi API pipelines and MCP servers, arbitrage detection, and calibrated-probability tooling — where models meet money and get graded on it.

You get — probabilities with a P&L attached.

Explore quant tooling →

The agent loop

Sense, reason, act, learn

Every agent I ship runs the same loop, and every step of it is measured. Agents, with adult supervision.

01

Sense

Signals in: your systems exposed as MCP tools, plus events, queues, and documents. The agent sees exactly what you let it see — scoped, logged, revocable.

02

Reason

Models plus evals: the model proposes, the eval suite grades. Model selection is a measured decision, not a brand preference.

03

Act

Side-effects behind human gates: drafts, diffs, and proposals by default — writes only where you approved them, with an audit trail either way.

04

Learn

Memory plus monitoring: what worked lands in shared memory, what broke lands on the dashboards. The loop tightens every run.

Dogfood

Running in my own shop

The agentic stack I sell is the one I run. My multi-project operation is an agent fleet — twenty-plus Claude agent sessions coordinated like a small team.

Fig. 03 — the fleetown ops · live
20+ sessions · one memoryevery write verified
  • 20+ Claude agent sessions across project folders — each with its own playbooks and session rituals
  • A shared memory service (mem0) every agent reads and writes, so session 21 starts where session 20 stopped
  • Per-project dispatch channels — agent-to-agent mail, plus a session registry of who worked where
  • Self-verifying write tooling — every durable write is fsynced and re-read byte-for-byte before it counts

This site, its deploys, and the tooling behind it are run by that fleet. Dogfooding is the eval.

See the live practice →

Engagement models

Not one shape. Three ways in.

Every problem is different, but engagements come in three shapes — pick the one that fits, or start with a conversation and we’ll find it together.

engagement-α 01

Advisory — audits, eval reviews, second opinions

engagement-β 02

Build — agent pilots, MCP servers, scoped systems

engagement-γ 03

Retainer — operate the fleet, watch the dashboards

20yr
Production engineering
99.99%
Uptime on multi-region AWS
~17
Prediction-market repos on GitHub
15
Certifications, most from the last year

“An agent without evals is a liability with an API key.”

Ionel Roiban · working principle

Have a workflow that should run itself?

Tell me what you want an agent to do. First conversation is free.

Book a consult