Discuss a program ↗

Verron.AI Technology stack

What we build on.

Frontier and open-weight models, fine-tuning, private and public runtimes, hybrid routing, GPU capacity. We work across all of it because the constraint decides — residency, latency, cost, audit — not a vendor relationship.

We sell no licences and no capacity. Recommendations start from your operating reality, not a platform sales target.

Selection
Driven by the constraint, documented in the decision memo
Independence
No resale, no capacity margin
Exit
Every route we set up, you can run or replace without us
01

Models

  • OpenAIOpenAI
  • Anthropic ClaudeAnthropic
  • QwenQwen

Frontier APIs where capability decides, open weights where residency, cost or offline operation decides. Routing is an explicit design choice, written down and reversible — not a default inherited from whoever sold the first pilot.

02

Fine-tuning and training

  • PyTorch
  • Hugging Face
  • TRL
  • PEFT · LoRA
  • QLoRA
  • MLX

Supervised fine-tuning and preference tuning on open weights, LoRA and QLoRA adapters, judge-gated evaluation before and after. When latency is the binding constraint we train speculative-decoding drafters rather than buying a bigger model. And we say so when tuning is not the answer — which is more often than vendors suggest: prompt, retrieval and routing usually move the number first, for a fraction of the cost.

03

Serving and runtime

  • vLLM
  • OllamaOllama
  • llama.cpp
  • DockerDocker
  • KubernetesKubernetes
  • Red Hat OpenShiftOpenShift

Private inference served on your own hardware, workloads isolated in digest-pinned containers, orchestration that a platform team can operate after we leave. This is where most AI pilots fail to become systems.

04

Hybrid routing

Private GPU
Your own hardware. Nothing leaves the perimeter, capacity is yours to schedule.
Local inference
Open weights on the machine that needs them, including offline and air-gapped operation.
Enterprise cloud
Managed model services inside your existing tenancy, region and contracts.
Frontier API
Used where capability genuinely decides, and only for the data allowed to go there.

One workload, several routes, chosen per call rather than per contract. The rule that decides is written down, testable and reversible — so a residency constraint, a price change or a model deprecation becomes a configuration decision instead of a rebuild.

05

Cloud and infrastructure

  • Microsoft AzureMicrosoft Azure
  • Amazon Web ServicesAWS
  • OpenStackOpenStack
  • VMwareVMware
  • ProxmoxProxmox

Public cloud, private cloud and on-premise, including identity and access surfaces. We have built and attacked access-governance systems on Azure RBAC and Entra ID, so the control plane is not theory to us.

06

Accelerators

  • NVIDIANVIDIA

GPU capacity sizing, scheduling and monitoring, from a single workstation to shared cluster capacity. The question we start from is how much you actually need, which is usually less than the first quote.

07

About this page

A stack, not an alliance.

What this lists
Technologies we design, build, operate and attack in client engagements.
What it does not claim
No partnership, certification, sponsorship or endorsement by any company named here. Where a formal partner status exists, it is named as such or produced in diligence.
Trademarks
All names and marks belong to their respective owners and are used here to describe our practice.

Stack decision

Bring the constraint, not the shortlist.

Residency, latency, audit, budget, the skills of the team who will run it. Tell us the constraint and we will tell you what the stack should be — including when the answer is that you do not need a model at all.

Request a stack review↗

Or simply contact@verron.ai