Private intelligence

Private, offline AI you own.

FlorenciaAI is a self-hosted assistant that runs on hardware you own — fully offline, on open-weight models, with no dependence on a cloud frontier model. It is being built to become a private network for the robots that will live with you.

Deployment
On-premises, private
Models
Open-weight, local
Status
Working prototype
01

The problem is dependence.

Most assistants today live in someone else’s cloud. The conversation, the memory, the tools, and — before long — the commands that move a robot all travel to a frontier model you do not operate and cannot keep.

That arrangement is convenient. It is a poor foundation for a system meant to stay in the house, remember what matters, and eventually coordinate machines on a private network.

FlorenciaAI starts from a different premise: the intelligence should belong to the person who uses it, run on hardware they own, and keep working when the outside world is not invited in.

Cloud dependence

Capability is rented from a provider. If the service changes, goes dark, or raises the price, the assistant goes with it.

Privacy leaves the room

Mail, research, household context, and future robot control are too intimate to hand to a model running elsewhere by default.

No private network

A house, a workshop, and a set of robots need a local mind — not a session in a browser tab on someone else’s GPU.

02

A hive you can keep.

FlorenciaAI is a self-hosted agentic assistant, and the beginning of a private robot network — a hive mind that stays on your side of the wall. It is not a cloud service, and it is not generally available as a commercial product today. The founder’s private deployment is already running.

Assistant

Agentic, local

Tool use, long-term memory, email, research, browser automation, and bot orchestration — designed to run entirely on owned hardware.

Offline

Open-weight only

The product runtime is local open-weight models. There is no dependence on a cloud frontier model for the assistant to think or act.

Later

A private hive

The same design extends to coordinating personal robots on a private network. That stage is planned, not shipping.

03

What’s running today.

A dedicated workstation named FloraAI, running Windows 11, already serves a 70-billion-parameter open-weight model on two NVIDIA GPUs, with a local interface, image and video tools, and a daily on-premises refresh pipeline.

Hardware
Two NVIDIA GeForce RTX 3090 Ti GPUs (Founders Edition and EVGA FTW3), 48 GB of VRAM in total, PCIe 4.0 x8 each on a Z790 board, without NVLink. The 70B model is split across both cards.
Primary model
Meta Llama 3.3 70B Instruct, served as GGUF by Ollama. Custom FloraAI builds include 32k-context and 65,536-context variants.
Interface
Open WebUI on the local network and on a phone.
Generation
ComfyUI on the same GPUs for image and video, including image-to-video and text-to-video tools.
Agents
Custom Ollama Modelfiles with agentic tools, including web search and ComfyUI-backed generation.

FloraAI-Business

A local RAG business assistant, one of the custom FloraAI model builds, refreshed by a daily automated pipeline.

Daily pipeline

A Kasa smart plug wakes the FloraAI PC. The machine runs its refresh and backup, then shuts down.

Also installed

Additional open-weight 70B and 103B models are installed on the same workstation for evaluation.

Generation speed for the 70B model has not been measured and is not claimed here. This is a private, working prototype — not a public product launch.

04

How the system is arranged.

Current private deployment. Planned serving and robotics paths are on the roadmap, not in this diagram.

05

Where it goes next.

The path is to deepen the local hive, then give it better serving and more capable on-premises NVIDIA hardware, and then extend the same offline design to personal robots. Everything below is planned. Nothing here is a ship date or a purchase commitment.

Planned

A stronger local hive

Richer memory, email, research, browser automation, and bot orchestration — still fully on-premises, still open-weight.

Planned

NVLink and more VRAM

Relieve the current two-card, x8, no-NVLink split so larger context, higher precision, and concurrent tools have room to run.

Planned

RTX PRO 6000 Blackwell

A 96 GB enterprise GPU, used only as internal FlorenciaAI development and deployment hardware if it is acquired.

Planned

TensorRT-LLM, vLLM, and NIM

Move serving from GGUF via Ollama toward NVIDIA TensorRT-LLM, vLLM, and self-hosted NIM on owned GPUs — not hosted cloud NIM as a runtime dependency.

Planned

Robotics integration

Extend the private network to personal robots, using NVIDIA Jetson and Isaac where they fit an on-premises, offline design.

06

Who is building it.

Jensen Florencia

Founder

Jensen designs and builds FlorenciaAI from Jefferson, Massachusetts. The current FloraAI workstation, the 70B deployment, and the agent tools are his work. This is a founder-led prototype — a team of one.

Contact

Write to the founder