Operator // JV — Private AI Deployment Security
Advisory for CISOs, founders, and technical executives moving AI onto their own infrastructure — where the vendor safety net disappears and the threat model quietly changes.
§01 — The Problem
Moving models inside your perimeter removes one class of risk and quietly introduces five more. The stack is new. The tooling is immature. And most deployments are configured by teams doing it for the first time, under deadline, with no one checking the work.
These are the failure modes I find most often. None of them announce themselves.
Exposed endpoints
Ollama, vLLM, and llama.cpp servers listening on 0.0.0.0 with no authentication. Findable in minutes. Regularly found.
Leaked retrieval corpora
RAG pipelines that index your most sensitive documents — then serve them to anyone who asks well.
Mishandled model artifacts
Unsigned weights from public hubs. Unverified checkpoints. Serialization formats that execute code on load.
Wrong hardware
Six figures of GPU purchased against a workload nobody profiled. The invoice arrives before the truth does.
Evidence gaps
No logs, no lineage, no answer — when the auditor arrives, or the incident does.
§02 — Who I Help
Executives
Who signed off on private AI and now own the risk.
Security leaders
Handed an AI stack their existing controls were never designed to see.
Founders
Whose enterprise deals now hinge on proving the deployment is safe.
Technical teams
Strong engineers, first AI deployment. They need a second set of eyes, not a lecture.
§03 — Services
01 / Advisory
$750–$3,500
A second set of eyes before you sign. I review the vendor claims, the hardware quote, and the architecture — and tell you what I would do.
02 / Sprint
$8,500–$18,500
From pilot to production without the guesswork. Two to three weeks, ending in a clear go / no-go and a hardening plan your team can execute.
03 / Assessment
$12,500–$30,000
Find what your deployment already exposes — before someone else does. Endpoints, retrieval corpora, and model artifacts, examined with evidence attached.
04 / Retainer
$5,000–$15,000/mo
Standing security leadership for your AI program — without the executive hire. I own the roadmap, the reviews, and the answer when the board asks.
§04 — Why Me
I'm a co-founder of Qompute AI, where we deploy and secure private AI infrastructure for real workloads — not slideware. The tools I ship and the research I publish come from the same deployments I advise on. When I tell you a configuration will fail, it's because I've watched it fail.
You get an operator's answer: specific, testable, and signed.
Co-founder
Qompute AI
Tools
Shipped open-source
Research
Published findings
Hands-on
Deployments reviewed
§05 — Tools & Research
The Hardening Checklist — free, no email
40 checks across 6 domains for Ollama, vLLM, llama.cpp, and RAG stacks. Printable. Share it.
Open-source exposure scanner
Point it at your own perimeter. Find the inference endpoints you forgot about.
iOS app
Private AI in your pocket, built with the same discipline this site preaches.
The Private AI Exposure Index
Ongoing research measuring how exposed real-world private AI deployments are.
AI Security Foundations — free course
The baseline every team should clear before a single weight is loaded.
§06 — Contact
Thirty minutes. Bring your architecture, your quote, or your doubt. You'll leave with a straight answer about where you stand — whether or not we work together.
Use the form → · or email joey.victorino@gmail.com
PGP available on request · NDAs welcome