You set the boundaries
Choose the environment, approved data,
users, and model.
Refactr
Put private AI to work.
Refactr deploys private AI assistants inside your infrastructure. Use the hardware you own, or let us source the system you need.
Your data is too valuable to send away.
Your team wants to use AI. Your company cannot send sensitive prompts, documents, or outputs to public AI services. That leaves employees working around the policy, or without AI at all.
Bring AI to your data.
Refactr deploys the right model on infrastructure you control. Your team gets fast AI for everyday work, while your prompts, documents, and outputs stay inside your environment.
What we deploy
Private AI for work your team already does.
Start with a document or coding assistant, deployed inside your own environment. More workflows follow the same path.
Document assistant
Ask questions, summarize material, draft from documents, and work through sensitive information without sending it to a public AI service.
Summarize the open decisions
Your company stays in control.
You decide where the system runs, which model it uses, and what data it can access. Refactr deploys and maintains the stack.
Work stays inside
In a private deployment, company work stays inside the environment you choose.
Choose or fine-tune the model
Deploy a compatible open model as-is, or fine-tune one for your workflow and approved data.
Use or source the hardware
Start with the GPUs you have. If more is needed, we can specify, procure, and supply it.
Refactr runs the stack
We handle setup, monitoring, updates, performance tuning, and ongoing support.
How Refactr runs the stack
Get more from local GPUs.
We built Epsilon, our inference layer, to route requests across local engines. It improves response time and hardware use without sending work outside your environment.
With Epsilon vs Basic load-based routing
TTFT p95
Tail time to first token on the same dual warm pool.
With Epsilon
0 ms
Without Epsilon
0 ms
0.0% lower tail TTFT
GPU-s / valid
Active GPU-seconds per valid request on the same pool and trace.
With Epsilon
0.000
Without Epsilon
0.000
0.0% lower GPU-s per valid request
~0% cheaper per million output tokens
With Epsilon costs about 0% less per million output tokens than Without Epsilon at every GPU rate we modeled. Throughput on the same pool is 0.0% higher. Dollar rates below are approximate scenarios, not invoices.
Where Epsilon fits.
Epsilon is built for efficient local inference on infrastructure you control. Other tools serve different operating models.
Swipe to compare systems
Axis
Epsilon
LiteLLM
llm-d
Dynamo
Routes across local GPU engines
YesWarm llama.cpp + vLLM
NoProvider / API backends
PartialvLLM-centered
PartialNVIDIA engines only
No cluster required
YesBare-metal / small host
YesProxy process
NoCluster control plane
NoDC orchestrator
Benchmarks configs before routing
PartialMeasured curves + policy
NoNo recipe gate
NoNo recipe gate
NoNo recipe gate
Single consumer GPU / small fleet
YesBuilt for one box to small fleet
PartialWorks; not fleet-native
NoCluster-scale fleets
NoNVIDIA datacenter
A recipe is the engine, model, and quantization each worker runs.
Product fit from public docs. Latency on this hardware was measured only for Epsilon vs basic load-based routing above.
Same two-GPU pool (2× RX 7900 XTX).
How we start
Start with one real job.
Don’t roll AI out across the company on day one. Pick one recurring task, put it in front of the people who do it, and see if it earns a wider rollout.
Find the job
Choose work people repeat often enough for an improvement to matter.
Draw the line
Decide what it can see, what it cannot, and who can use it.
Put it in people's hands
We set it up in your environment and support a small group as they start using it.
Decide what to do next
If it is useful, expand it. If it needs work, improve it. If it is not worth it, stop.