Ionel Roiban

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 cloud and web engineering at the bottom — delivered by the person you actually talk to.

What I do

Six practices, one engineer

Capabilities offered as services. Where a number appears, it comes from a real, named engagement — not a projection.

Agentic AI & MCP

Agent fleets with shared memory, MCP servers that expose your internal systems as safe, scoped tools, human-approval gates on every side-effect, and session orchestration. I run my own multi-project operation this way — twenty-plus agent sessions, one shared memory — so the patterns arrive already broken in.

  • Agent fleets with shared memory and dispatch channels
  • MCP servers: your systems as agent tools, scoped and logged
  • Human-approval gates on anything that writes
  • Session orchestration, registries, and playbooks
  • Proof point: my own ops run this way, daily — dogfood, not demo

LLM Systems

RAG over private knowledge bases, local-first deployment with Ollama and llama.cpp, model selection driven by evals, and context engineering that keeps answers grounded in your data — not the model’s imagination. Your data never has to leave your hardware.

  • RAG over private knowledge bases
  • Local-first deployment (Ollama, llama.cpp)
  • Model selection & eval pipelines — evals before vibes
  • Context engineering for grounded answers

Data Science & Analytics

Calibrated forecasting, customer-behavior analytics, BI dashboards in Looker, Tableau, and Grafana, automated alerting, and evals & metrics design for AI systems — on a certified data stack.

  • Calibrated forecasting & probability estimates
  • Behavior analytics for subscription retention
  • Looker / Tableau / Grafana dashboards + automated alerts
  • Evals & metrics design for AI systems
  • Proof point: 3 Cisco data certifications + Palantir Foundry & AIP

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 quietly rotting between runs.

  • n8n pipelines with LLM steps across APIs and side-effects
  • Self-hosted Mautic + AI — a rebuilt email stack measured at 31% opens, 4.2% CTR
  • Process monitoring (monit) and alerting
  • The operational glue nobody budgets for

Cloud & MLOps Architecture

AWS multi-region delivered at a documented 99.99% uptime across US-East and US-West, HIPAA- and SOC II-shaped designs with audit trails, blue-green deploys with DNS failover, and model-serving infrastructure — with twenty years of web platforms underneath, including page loads taken from 8+ seconds to under 2.

  • Multi-region AWS at 99.99% documented uptime
  • HIPAA & SOC II-shaped designs, audit trails
  • Blue-green deploys with DNS failover
  • ~40% infrastructure cost reduction on migration
  • Model-serving infra for local and hosted LLMs

Prediction Markets & Quant

Kalshi API pipelines and MCP servers, arbitrage detection, and calibrated-probability tooling — where models meet money and get graded in public. Built and run as ongoing personal research, available as an engagement.

  • Kalshi API pipelines & arbitrage detection
  • MCP servers for model-market integration
  • Calibrated-probability tooling
  • Proof point: ~17 public prediction-market repos on GitHub

Who I help

Teams that need it to actually ship

From regulated pharma platforms to streaming media to quant research — the common thread is an AI system that has to hold up in production, not a demo.

Teams that want agents, not chatbots

You’ve done the copilot pilot. Now you want an agent that actually files the ticket, updates the CRM, and reports back — behind approval gates your compliance team can sign off on.

Founders & small teams

You have an idea, a pile of APIs, and no appetite for a department. One senior person designs the agent, builds it, runs it — and hands you the keys, prompts and evals included.

Regulated industries

Healthcare, pharma, finance. I built a pharmaceutical B2B platform with full audit trails for controlled-substance orders — the same discipline now decides what an agent may touch.

Publishers & media

From a financial publisher’s digital transformation to NBA streaming on Roku — now with content pipelines that carry LLM steps and dashboards on what readers actually do.

Marketing & growth teams

Self-hosted Mautic with automated IP warming took a crisis list from 12% back to 31% opens and 4.2% CTR. Add LLM steps for segmentation and drafting — on sending reputation you own.

Subscription businesses

Customer-behavior analytics for retention: calibrated churn signals, dashboards your team actually reads, and alerts that fire before the quarterly numbers notice.

Quant & research desks

Prediction-market pipelines, MCP servers, and calibrated-probability tooling — for people who want models and markets talking, with the calibration curve to prove it.

Measured outcomes

Numbers from production, not projections

Two examples of before-and-after, measured on live systems.

Page loads — publisher platform

Digital transformation with a CloudFront CDN and automated testing, code review, and deploy pipelines.

before8+s
after<2s

Email deliverability — rebuilt infrastructure

A deliverability crisis fixed with self-hosted Mautic and automated IP warming.

crisis28% → 12%
after rebuild31% open · 4.2% CTR

Full context for each figure on the case studies page.

How I build

Opinions, earned the slow way

Current focus: agent fleets, MCP servers, eval pipelines, local-first LLMs, n8n automation with LLM steps, and prediction-market tooling.

Evals before vibes

Model choices, prompts, and agent behavior get graded by eval suites, not adjectives. If quality isn’t measured, it’s a mood.

Human gates on side-effects

Agents propose; approved actions execute; refusals get logged. Nothing writes to production without a witness.

Local-first by default

Ollama and llama.cpp before API bills. Small-but-good models on hardware you control — cheaper, private, yours to keep.

Monitored & auditable

Dashboards and alerts with the first deploy, audit trails by design. If a regulator asks how it works, there is a document.

Start here

Tell me what you’re trying to do

{Agent pilot} {MCP server} {Dashboards}
$ssh roiban.com "describe the workflow you wish ran itself"⧉ Copy
>mail -s "here is what I am trying to do" [email protected]⧉ Copy

Have a system that should exist but doesn’t?

Tell me what you’re trying to do. First conversation is free.