RelayAI that runs on hardware you own

Make your company AI native.

One agent your whole team works with. Same tools, same context, same knowledge as your people, on private models you control. Built around your workflows by our engineers, live in days, not months.

The live portfolio. A board deck built from one sentence. The inbox on the CEO's phone.

01The problem

Your company already knows everything. Nobody can see it.

02Meet Relay

One agent the whole company works with.

Same tools, same context, same knowledge as your people. Relay plans each request, does the work, and asks a person when it should.

relay.familyoffice.internal/runs

03Works inside your team

It asks the right person. Then it gets on with it.

relay.vistahomes.internal/c/collections

04What Relay does

Five functions. One agent that knows them all.

Every client gets a fast, informed answer.

  • Reads each client email and message, then drafts the reply in your voice
  • Keeps one live record of every client, request and promise made
  • Prepares meeting briefs and follow-ups from past conversations
  • Holds anything sensitive until the right person approves

Asks the client owner before sending

  1. Client Relations
  2. Finance
  3. Accounting
  4. HR
  5. Legal
relay.yourcompany.internal/client-relations

05Relay Router

Nobody on your team should have to pick a model. Relay picks one for every prompt.

Most people use one model for everything. They overpay for easy work and get weak answers on hard work. The router reads each request, and each event like a new email or a schedule, then sends it to the right model. Sensitive data stays on your hardware.

06Connect everything you already use

Messy files in. Knowledge out.

Relay connects to the tools you already run and reads the files your people keep. Nothing has to be replaced on day one.

relay.familyoffice.internal/knowledge

07Everyone can build

Dashboards, apps, decks and documents. Just ask.

Every employee can build what they need from the company's own data. No tickets to IT, no waiting on an analyst.

relay.vistahomes.internal/studio

08Private AI, on your desk

Hire AI that lives in your office. It gets smarter every year.

Buy the machine once. It works for you for years. Every time a better open model comes out, we install it on the same machine, so your team gets smarter without a new bill.

  1. 01Office mini PCfrom $3,500
  2. 02AMD workstationfrom $5,200
  3. 03NVIDIA DGX Sparkfrom $6,950
  4. 04NVIDIA DGX Stationfrom $105,000
Office mini PC

One time, from

$3,500

  • AMD Ryzen AI Max+ 395
  • 128 GB unified memory
  • Quiet, about 140 W
  • Up to 25 people

Runs open models

Qwen3.8-Flash-Nextgpt-oss-120bGemma 4

09Open models caught up

Models you can own now do real office work as well as the best.

GDPval tests AI on 220 real tasks from 44 jobs: reports, slides and spreadsheets. Experts compare the results without knowing which model made them. The best open model is 8 points behind the top.

Most daily work does not need the very top model. For the rare hard task, and as a backup, Relay can use a frontier model through a zero data retention account.

Open weights, runs on your hardware Closed, cloud only

GDPval-AA · Elo · higher is better

Nov 2026

  1. Claude Opus 5.5

    Anthropic

    1866

  2. Kimi K3.1 Pro

    Moonshot AI· open

    1858

  3. Claude Sonnet 5.5

    Anthropic

    1838

  4. DeepSeek V4.1 Flash

    DeepSeek· open

    1746

  5. Gemini 4 Carbon

    Google

    1733

  6. Grok 4.7

    xAI

    1718

  7. MiMo-V2.6-Pro

    Xiaomi· open

    1689

  8. GLM-5.3-Flash

    Z.ai· open

    1649

  9. Qwen3.8-Flash-Next

    Alibaba· open

    1636

  10. Claude Haiku 5.5

    Anthropic

    1631

Source: Artificial Analysis, GDPval-AA v2.1, checked 9 November 2026. Selected models.

10Buy, don't rent

Agents work all day. Renting them by the token adds up.

A chat is a few questions a day. An agent reads, calls tools and checks its own work, all day. As it takes on more of your work it needs more compute, and a metered bill grows with it. Hardware you own costs the same whether it works one hour or twenty four.

Revolut, October 2026

Moved about 2,000 engineers to self hosted open models and cut its AI coding bill by about 90%.

Compute cost over 3 years · 40 people

Renting by the token Owning the hardware

Saved by month 36

$0k

$0$100k$200k$300k$400kDay 1M6M12M18M24M30M36Paid back in month 6More hardware added$357k$39k

Worked example. Renting: $3 per million tokens, agents taking on 7% more work each month. Owning: two DGX Spark on day one, two more at month 18, power included. Your numbers depend on your work, so we size it with you.

11Continual learning

It updates itself and gets better every week.

Every answer, correction and approval teaches Relay how your company works. Each night it reviews its own runs and rewrites the playbooks that needed fixing. When a better open model comes out, it is tested on your work before it is switched in, so the same machine keeps getting smarter.

WorkFeedbackMemorySkills

What Relay learned

This week · Vista Homes

Learning

0

skills learned

0

playbooks updated

Corrections per 100 tasks

−76%

  1. Relay · Nightly review

    Models

    A new open model was released

    Tested overnight on 400 of your past tasks. Switched for extraction, 18% fewer corrections.

  2. Priya Menon · CFO

    Skills

    Asked for the review in her format, three times

    Saved as a skill: Quarterly review, Priya's format. Now one sentence away.

  3. Kabir Shah · Legal

    Legal

    Rewrote clause 7.2 in a sale agreement

    Template updated with his wording. Used in 12 new agreements.

  4. Anita Rao · Finance

    Accounting

    Changed the GST rate on two vendor bills

    Rule added to the Accounting playbook. Applied to 46 bills since.

12Privacy and compliance

Your data never leaves the building.

Relay runs on hardware in your office, with open models on that same hardware. Your prompts, files and answers are never sent to an AI company.

Network boundary

Relay · head office

Connected · allowlist

Sent to AI companies, last 30 days

0bytes

Requests answered on site

0

Designed to supportGDPRSOC 2 controlsISO 27001 controlsHIPAADPDP Act

13Privacy and governance

Built in from day one. Not added later.

14Why not just buy ChatGPT or Claude

They hand you a tool and hope you use it well. Knowledge workers are not engineers.

Getting real work out of AI takes agentic engineering: the right tools, context, models and checks around every task. That is our job, so your people can focus on the work that matters.

A general AI subscription

  • A chat box and a login for each seat
  • Each person writes their own prompts
  • Pick a model and hope it was the right one
  • Your data goes to their cloud
  • Starts from zero in every chat

Relay, deployed by Zosma

Arjun NayakPranay JainShanvit ShettyDevendra Mishra
  • Engineers who sit with your teams and map how work flows
  • One agent that already knows your tools, files and rules
  • The Relay Router picks the right model for every request
  • Runs on hardware you own, inside your office
  • Learns from every answer, correction and approval

15Forward deployed engineers

Live in days, not months.

Our engineers work closely with your teams. They learn your workflows, build the harness around them and stay until the return is clear. Then they keep finding the next thing worth handing to an agent.

  1. Map the work

    We sit with each team and write down how work really flows, step by step.

  2. Install

    Hardware in your office. Open models, connectors and sign in set up.

  3. First agents live

    The first workflows run in Relay, with your people approving.

  4. Whole team on board

    Everyone works with Relay. We train each team on its own work.

  5. More work, measured

    More tasks move to agents. We report hours saved and return.

16Built in the open

Engineers at the world's best software companies star our open source work.

Relay's memory is built on pi-llm-wiki, our open source knowledge base that keeps itself up to date. These are companies listed on the GitHub profiles of people who starred it.

zosmaai/pi-llm-wiki0stars54forks0downloads a month
  • Microsoft
  • Datadog
  • IBM
  • Shopify
  • Cloudflare
  • Atlassian
  • Grafana Labs
  • Palo Alto Networks
  • Alibaba
  • ByteDance
  • TikTok
  • Meituan
  • Comcast
  • Michelin
  • Optum
  • Delivery Hero
  • Zoom
  • Gusto
  • Hitachi
  • Qihoo 360

Backed by

NVIDIA Inception Program
Google for Startups
  • Family office

    Aldermere Family Office

    Monthly portfolio review that used to take days now arrives ready to approve.

    • Reads the monthly portfolio workbook and checks every sheet before anyone sees it
    • Operators upload and fix. Reviewers approve and publish.
    • Tracks strategic initiatives next to the numbers
    • Signs in with the team's own Microsoft accounts
  • Family office

    Tamarisk Partners

    One shared workspace where the team and the AI work on the same facts.

    • One workspace for initiatives, goals and the questions the team asks
    • Each person sees only what their role allows
    • Runs on private models with zero data retention
    • Chat history can be removed with a clear audit trail
  • Fund management

    Corvane Funds

    Holdings and ledgers flow in from the portfolio system without manual exports.

    • Pulls holdings, ledgers and sellable positions from the portfolio system
    • Catches duplicate rows and bad dates before they reach a report
    • Turns the firm's own processes into a library the agent can follow
  • Housing developer

    Larkfield Housing

    Bookings, receipts and handovers in one place instead of many spreadsheets.

    • One database replaces duplicated Excel sheets as the system of record
    • Sales, admin and owners each get an agent that knows their part of the business
    • Follows a flat from booking to demand, receipts, parking, handover and registration
    • Keeps the working automations running while the data foundation is built

Make your company AI native

Your people, plus an agent that knows the whole company.

Private models on hardware you own. Relay deployed around your workflows. Engineers who stay until it pays for itself.

  1. 01

    A 30 minute call

    Tell us how your teams work today.

  2. 02

    A workflow map

    We show where agents return the most, first.

  3. 03

    Live in days

    Hardware, models and Relay in your office.