IRIS8 Talk to us

Custom AI models, built and run for you

AI you own,
not rent.

IRIS8 trains a model on your data and runs it for you, inside your own environment. You need no ML team.

Model cardSample data
Model
loan-file-qa v3
Owner
You
Trained on
Your loan files
Runs in
Your VPC
Run by
IRIS8
Status
In production
  • Already running inside a customer's own cloud
  • NVIDIA Inception member
  • Backed by FirstCapital.vc
  • VPC or on-prem

Rent vs own

Renting and owning, side by side

A year of API calls leaves you with invoices. A year with your own model leaves you with the model and its records.

Rented

You keep nothing

Yours

Owned

Weights v1
Weights v2
Weights v3
Datasets
Benchmarks
Records

You keep the model

Cost
Rises with every call
The hardware it runs on
Changes
On the provider's schedule
When you approve them
What builds up
Paid invoices
Versions, datasets and records

Cost calculator

Run your own numbers

API bills grow faster than the plan. Every figure here is yours to change.

Inputs · example values, change them
Example
Example
Example
Example
Infrastructure plus service
Example

API only

$50,000

$0.0100 per call · per month

With your own model

$30,000

$0.0060 per call · per month

At these inputs your own model costs $20,000 less per month.

BREAK-EVEN VOLUME YOUR VOLUME TOKENS PER MONTH COST PER MONTH
API onlyYour own modelYour volume
API only= tokens ÷ 1M × API price
Own model= running cost + (1 − share) × tokens ÷ 1M × API price
Cost per call= monthly cost ÷ (tokens ÷ tokens per call)
Break-even= running cost ÷ (share × API price) × 1M tokens

Example values, not a quote. At low volume the API can be cheaper.

Go through your numbers with us

How it works

The team that trains it runs it

Your team gets the same console our engineers work in.

Our AI research engineers work alongside your team at every stage. IRIS8 research engineerYour team

IRIS8 console · loan-file-qa / benchmark Sample data

Benchmark · loan file QA

Test set
500 held-out cases
Frontier model94
Open base model71
Your model v395Passed
Signed off by
Head of Credit Operations

Gateway · live requests

  • Summarise loan fileyour model
  • Extract KYC fieldsyour model
  • Repeat questioncache
  • Draft credit memofrontier model

Account number masked before sendEvery request logged

Model record · loan-file-qa v3

Base model
Open-weight, licence checked
Training data
12,400 examples, personal data masked
Trained
14 Oct 2026
Benchmark
95 against a pass mark of 92
Approved by
Head of Credit Operations
Deployed
Your VPC, 21 Oct 2026
Access
IRIS8 engineer, scoped, logged
Previous version
v2, kept for rollback

Scored against the frontier model on your own cases.

See the managed service

Use cases

Where a pilot can start

These are starting points for a pilot, grouped by the team that owns the work.

Lending and banking

  • Commercial loan intake
  • Credit memo drafts with source citations
  • Financial spreading into your own template
  • KYC and account-opening document packs
More for banking
Commercial loan intakeSample data
  1. 01

    Credit package in

  2. 02

    Documents sorted against the checklist

  3. 03

    Tax return missing, flagged

  4. 04

    Sent to underwriter

Insurance

  • Claims intake and triage
  • Claims file summaries
  • Policy document extraction
  • Underwriting submission prep
More for insurance
Claims intake and triageSample data
  1. 01

    First notice of loss in

  2. 02

    Claim type and details extracted

  3. 03

    Police report missing, flagged

  4. 04

    Routed to adjuster

Healthcare

  • Prior-authorization packs
  • Clinical note summaries
  • Coding and documentation support
  • Patient message triage
More for healthcare
Prior-authorization packsSample data
  1. 01

    Chart notes in

  2. 02

    Request form filled

  3. 03

    Lab result missing, flagged

  4. 04

    Sent to staff for review

Operations

  • Forms and files into structured records
  • Account-opening document packs
  • Policy and procedure questions, answered from your documents
  • Routine case handling from your own procedures
All industries
Forms and files into structured recordsSample data
  1. 01

    Application and files in

  2. 02

    Fields read into one record

  3. 03

    Signature missing, flagged

  4. 04

    Sent to operations

Credit and underwriting

  • Credit memo drafts with source citations
  • Commercial loan intake
  • Financial spreading into your own template
  • Underwriting submission prep
All industries
Credit memo drafts with source citationsSample data
  1. 01

    Financials and source files in

  2. 02

    Memo drafted

  3. 03

    Each figure cited to its page

  4. 04

    Sent to underwriter

Risk and audit

  • Fraud case summaries with the evidence attached
  • KYC document completeness checks
  • AML alert triage
  • Policy lookups with the source cited
All industries
Fraud case summaries with the evidence attachedSample data
  1. 01

    Fraud case opened

  2. 02

    Transactions and notes pulled

  3. 03

    Case summary drafted

  4. 04

    Sent to investigator

Customer service

  • Support answers from your own policies
  • Customer correspondence drafts
  • Contact-centre assist for agents
  • Patient and customer message triage
All industries
Support answers from your own policiesSample data
  1. 01

    Customer question in

  2. 02

    Answer drafted from your policy

  3. 03

    Policy clause cited

  4. 04

    Exception passed to an agent

Legal

  • Contract review with clause citations
  • Key points extracted from memos
All industries
Contract review with clause citationsSample data
  1. 01

    Contract in

  2. 02

    Key clauses extracted

  3. 03

    Each clause cited to its page

  4. 04

    Sent to counsel

Engineering and security

  • Pull request review on your own codebase
  • Security alert investigation notes
  • Incident evidence pulled into one summary
All industries
Pull request review on your own codebaseSample data
  1. 01

    Pull request opened

  2. 02

    Changes read against your codebase

  3. 03

    Risky change flagged

  4. 04

    Sent to reviewer

For your security review

What your security team can check

Each point below is something your team can verify inside your own environment.

  • Data stays in your VPC or on-prem
  • IRIS8 engineers get scoped, task-bound access
  • Every engineer action is on the audit record
  • If you leave, models and records stay yours
See how we operate
Access and audit log · loan-file-qaLog kept in your environmentSample data
  1. 09:12IRIS8 engineerAccess granted: retrainScoped to task
  2. 09:14IRIS8 engineerTraining run startedLogged
  3. 10:41IRIS8 engineerBenchmark runLogged
  4. 10:58Head of Credit OperationsVersion v4 approvedLogged
  5. 11:03IRIS8 engineerSession closedLogged
  6. 11:20GatewayAccount number maskedLogged

If you leave: weights, datasets, benchmarks and records stay with you

Your stack

Works with what you already run

Your model sits behind an OpenAI-compatible API, so existing apps connect with a config change.

Your apps and frameworks
  • LangGraph
  • CrewAI
  • AutoGen
  • MCP servers
  • Python
  • Node
EndpointOpenAI-compatible APIChange the base URL
Models
  • Open-weight base model, licence checked
  • Your fine-tuned model
  • Frontier model, optional
Runs on
  • Your cloud account
  • Your data centre
  • NVIDIA GPUs

Customers

Already running inside a customer's own cloud

Mid-size lender · loan file checks

Illustrative example

Situation

Checkers read every loan file by hand. Customer documents could not leave the lender.

What we built

A model fine-tuned on past files, running in the lender's own VPC.

What changed

Checkers now review the model's findings instead of reading each file end to end.

Illustrative example
“We wanted the model in our own account and someone to run it. Hiring an ML team for one model made no sense.”
CTO, mid-size lender
Illustrative example
“The benchmark report got this past our risk committee. It showed our own cases and the pass mark we had set.”
Head of AI, insurer

The Sovereign Pilot

Start with one workflow, in 8 weeks

A fixed-price pilot on one use case and one production KPI, credited to year one.

  1. 01

    Assess

    Agree the workflow, KPI and pass mark

  2. 02

    Build

    Train and benchmark in your environment

  3. In production03

    Run

    Go live and report the scores

You keep the weights and the benchmark report whether the model passes or fails.

Questions

Questions we get on the first call

Talk to us
Q1Do we need an ML team?

No. Your experts set the pass mark and review sample outputs, and we do the training, deployment and operations.

Q2Why not build this ourselves?

You can, if you are ready to hire ML engineers and keep them on call for the model. With IRIS8 you own the same model and records without that hiring.

Q3Which models do you use?

Licensed open-weight models, tested against your benchmark for each workflow. There is no lock-in to one model vendor, and frontier models stay available through your gateway.

Q4Where does our data go?

It stays in your environment when the model runs in your VPC or on-prem. If we manage the cloud for you, it stays in the region you choose.

Q5What if we want to leave?

You keep the weights, datasets, evaluations and records, and the model keeps running in your environment.

Q6How is it priced?

The pilot is fixed-price and credited to year one, and after that a managed-service fee covers operations and retraining. You pay for infrastructure directly or at cost.

Next step

Bring one workflow to a 30-minute call

We will look at your data, your environment and whether a pilot makes sense.