Security-first infrastructure for regulated engineering teams adopting AI.
Darren Kay, Founder
Sanitized brief — no architecture diagrams or threat-model artifacts.
“Adopt AI — but not like that.”
In regulated industries, the customer’s contract bans public cloud AI outright — so the work moves in-house, onto dedicated GPUs running local models. The same wall hits on two fronts: engineers who want AI to accelerate delivery, and teams trying to ship AI features to their own customers. Doing it the allowed way is punishing:
Developer laptops can’t mirror the production GPUs, so every change means rebuild → deploy to a test box → evaluate → repeat.
The GPUs you can actually get force model sharding and workarounds — more moving parts, less confidence.
AI cost is hard to forecast or plan around, so budgets and approvals stall.
The teams most eager for AI are the ones most constrained from using it safely.
The IP boundary and the software-delivery boundary are becoming the same boundary — and in regulated industries, it can’t run through someone else’s cloud. And a breach anywhere in a regulated sector is a near-miss everywhere in it: peer boards ask the same question — are our crown jewels actually segmented? — and move budget accordingly.
One dedicated environment where regulated teams can use AI — with code, secrets, and workflows kept inside a boundary they control.
Shipped as one coherent operating model — not five tools each customer has to integrate into a trust boundary themselves.
Everyone sells a piece. The customer is left integrating their own trust boundary.
Raw building blocks. You are the systems integrator, and the security model is yours to prove.
Isolation, but no opinion on secrets, workflows, or AI governance.
One slice of the problem — assembled alongside many others.
Developer velocity first; the trust boundary is an afterthought.
Assembling all of the above by hand — expensive, and rarely their core mandate.
The integrated operating model — dedicated boundary, brokered access, and evidence, coherent by design.
We are not “better hosting.” The operating model is the product.
Centralize the power. Thin the endpoints.
Distribute
- A powerful compute box per engineer, plus a second managed laptop
- Two devices to harden, patch, and explain — per person
- Attack surface and credential reach expand with every endpoint
- Higher endpoint cost, more to defend in an audit
Centralize
- One hardened environment holds the compute and the sensitive work
- A locked-down, browser-only client behind zero-trust access
- Fewer devices, less credential sprawl, a smaller and more explainable attack surface
- Lower endpoint cost, one story to defend
We don’t wait for a breach to find our blast radius.
With zero customers, we run adversarial tabletop exercises against real, in-the-wild supply-chain attacks — trace the blast radius as if we had been hit — and let that analysis drive the next control.
→ Workspaces hold no standing credentials to scrape
Secrets are brokered per task instead of pre-loaded, so a scraper finds nothing sitting there.
→ The pipeline’s privileged path is brokered, not standing
The blast radius of a poisoned build step is bounded by what was leased for that job.
One task, one governed lifecycle.
Each step is a control that removes a specific way IP or access leaks — described at the trust-boundary level, not the architecture.
Narrower than mass-market infrastructure — and far stickier.
· Stronger governance over code, secrets, and AI workflows
· A tighter, demonstrable trust boundary
· Reduced credential and device sprawl
· Predictable, right-sized cost — not unforecastable AI spend
· Operational evidence they can put in front of auditors and their own customers
· AI adoption is forcing every regulated org to answer “where does our IP live?”
· Event-driven demand — one breach in a sector puts crown-jewel protection on every peer board’s agenda
· A wedge with high willingness-to-pay and high switching cost
· Becomes part of the customer’s operational trust model — hard to rip out
· A specific, reachable beachhead, not a boil-the-ocean platform
We would rather under-claim and show our work.
- 0 standing secrets held by runners
- outbound-only control plane — no inbound admin ports
- signature-verified, digest-pinned releases
- bounded, deny-by-default task execution
- signed per-run evidence records
- append-only evidence retention
- SOC 2 attestation
- each with a written definition of “done”
Every public claim is backed by evidence we can produce on request — or marked as roadmap. For a security company, credibility is the product.
Targeted input at the decisions that are hardest to reverse.
A formal security advisory role.
Equity on standard advisory terms, calibrated to involvement. And a clean path that respects your obligations to the University of Florida and keeps the IP boundary unambiguous:
UF approval
Outside-activity / conflict-of-interest sign-off secured first.
NDA + agreement
Confidentiality and an advisory agreement — GeekVault IP stays GeekVault’s.
Deep review
Threat model, architecture, and roadmap — in a setting that protects us both.
Built nights and weekends over the past year by two engineers — and war-gamed against real attacks before the first customer.
Darren Kay · Founder