GemPhy
Specialty AI. Your Hardware. Your Data.
GemPhy installs a quantized language model on your own machine, grounds it in your own documents, and answers with citations back to the exact source — entirely on-premises. No cloud API, no data egress, no calibration data leaving the box.
One command. Runs locally.
The installer drops the GemPhy CLI on your machine and hands off to gemphy install, which fetches the engine, writes prompts and configs, and brings up the local stack — offline by default.
$ curl -fsSL https://gemphy.com/install.sh | shLinux / macOS / Windows · single-GPU workstation · nothing phones home
Runs where your data lives
Fully on-premises, offline by default. No phone-home, no cloud API in the loop. Built for organizations whose policy or regulator won't allow client material to reach OpenAI, Anthropic, or Gemini.
Bring your own open model
Compile an open-weights model from Hugging Face to GemPhy's 5-bit format and run it on a single workstation GPU. A 14B-class model fits in ~10 GB, a 7B in 5.4 GB, with task accuracy near FP16.
Grounded and cited
The model lives next to your regulations, contracts, and specs. It answers with verbatim citations from your own documents and refuses to fabricate section numbers — no fine-tuning, no data sent out for training.
Composable reasoning
Build domain workflows as a Thought-Process Graph, wired to GemPhy tools and your project context — then run them against any task in the same private deployment.
Reasoning that adapts to the problem
Adaptive cognition — the ability to flexibly adjust thinking, learning, and behavior to navigate new, changing, or uncertain environments.
The Thought-Process Graph is how you see and shape it. Compose the model's reasoning as an explicit graph — Observe, Classify, Hypothesize, Check, Decide, Loop, and Defer nodes. Each node runs against your project context and can call GemPhy tools, so you get an auditable trace with citations instead of an opaque answer.
It works with the rest of the stack — your grounded documents, your behavior rules, and your tools — all inside the same on-prem deployment.
Tools invoked
The graph is the first step toward a system that manages its own work — and learns from it.
Task registry
Every identified need becomes a tracked task with state across phases — not a transient chat.
Meta-cognition
The system decides which phases a task actually needs; a one-line job skips half of them.
Reusable skills
Successful implementations are promoted into reusable skills, so similar tasks short-circuit to an existing solution.
Human gates
Approval checkpoints at Requirements, Design, and Delivery — full autonomy in between.
Memory
Past projects inform estimates, design choices, and "we already built this."
Self-improvement
Post-mortems feed forward into the next task — and it can build itself new tools when it hits a class of work it can't yet handle.
The aim is cognitive cloning: capturing an expert's reasoning as a structured, reusable process — so the next similar task short-circuits to a solved one.
One private console for the whole stack
Chat with citations
Ask in plain language; every claim links back to the exact source in your documents.
Workspace selector
Pick the model, the document set, and the behavior rules; the engine loads them and serves everything locally.
Document ingest
Drop in PDFs, docs, and specs; GemPhy indexes them so the model can quote them directly.
Interface previews — live screenshots dropping soon
One built out, three in active design
One is built; the rest are being designed with partner customers.
DOT Compliance Assistant — 49 CFR Part 40
GemPhy's most developed vertical is a regulation-grounded assistant for trucking, aviation, transit, rail, and pipeline operators — and the TPAs, MROs, and SAPs who serve them. It cites 49 CFR Part 40 verbatim from the regulation text, is built to refuse fabricated section numbers, and runs entirely on the operator's own hardware. Talk to us about a pilot.
Healthcare Practices
The constraint
HIPAA limits what patient material can be sent to a third-party LLM. Most cloud AI workflows are a non-starter for IT review, and the practice wants the assistant inside its own four walls.
Where GemPhy fits
We're engaging design-partner practices to build the clinical-administrative deployment — note summarisation, dictation cleanup, document Q&A. Talk to us if you want to be one: pilot pricing, you keep the deployed model. GemPhy is software, not a clinical-decision device.
Law Firms
The constraint
Attorney-client privilege and the work-product doctrine make cloud APIs legally risky for any privileged material — discovery, deposition prep, contract review.
Where GemPhy fits
We're engaging design-partner firms to build the document-workflow deployment — review, markup, summarisation. Talk to us if you want to be one: pilot pricing, you keep the deployed model. GemPhy is software, not legal advice.
Family Offices & Private Wealth
The constraint
Contractual and reputational reasons to keep client financials off public clouds. A leak isn't a bug; it's an existential event for the relationship.
Where GemPhy fits
We're engaging design-partner offices to build the wealth- and estate-document deployment. Talk to us if you want to be one: pilot pricing, you keep the deployed model. GemPhy is software, not financial advice or a registered service.
Compliance posture: certifications such as a HIPAA Business Associate Agreement and SOC 2 are in active development. GemPhy is engaging with design-partner customers today; the customer remains the regulated party and operates the deployment inside its own program.
Talk to us about a deployment
You can install GemPhy yourself today. If you want help standing it up at your site — choosing a model, grounding it in your documents, and building the behavior rules and tools for your workflow — tell us about your environment and we'll be in touch within one business day.
Request a Conversation
GemPhy installs at your site. Tell us about your environment and we'll be in touch within one business day.
GemPhy is not built to win the throughput axis against cloud serving stacks. It's built for single-site, single-tenant deployment where intelligence per gigabyte and zero data egress are the buying criteria.
GemPhy is software. It is not medical advice, legal advice, financial advice, a regulated professional service, or a clinical-decision device. Customers in regulated industries operate the deployment inside their own compliance program.