Every industry breaks AI in a different place. Name the workflow. We set the route.

We deliver AI Data & Evaluation and AI Implementation around the workflow you run and the constraint it runs under, whatever the industry. Below, what we produce, industry by industry.

the three questions Data · System · Connected delivery

Your model passed the benchmark. Production asks three questions the benchmark never did.

The cabin, the ward and the counter are not in the training set. Every engagement here is specified against these three, and each is answered where the product runs.

Data Was the data recorded where the product runs? A corpus from a booth has never met the cabin. The model fails there first.
  • the cabin at 110 km/h
  • the ward round
  • your own applications
  • line 4, shift B
  • the citizen's own language
  • the course, in the learner's language
  • native speakers · 150+ languages
 System Does the system respect who decides? An answer from outside the approved sources is a control failure before it is a model failure.
  • the vehicle's HMI
  • the rights object, every turn
  • your policy and your thresholds
  • the MES work order
  • approved sources only
  • the rubric, then the educator
  • a person in the gate
 Connected delivery Was it proven in place before release? Without an evaluation set from production, every change to the model ships on impression.
  • evaluation sets from production
  • hard variants of your questions
  • acceptance criteria that are yours
  • a record you can inspect
  • 30-day erasure at contract end

Answer the three for your workflow and the route sets itself. Where one cannot be answered in place, we say so at scoping.

two lines, one team

Buy the data. Buy the system. Or both, from one team.

AI Data & Evaluation

→ question one

Collection, annotation, human review and evaluation sets recorded and graded where your product runs, in the languages it meets.

AI Data & Evaluation

AI Implementation

→ question two

Assistants, agents and document workflows built into the systems and controls already in place, with the record kept.

AI Implementation

Connected delivery

→ question three

One team owns the system and the data it runs on. When the failure sits in the data, the people who built the system fix it.

How the loop works
  1. 150+ languages
  2. 210 000+ contributors in the network
  3. 50+ countries
  4. EEA processing by default

Below, the industries we show in depth. Every other sector runs on the same two lines and the same delivery. Name yours in the probe and the page answers.

YPAI delivers for · automotive and mobility

  • Cerence AI
  • Hyundai
  • BYD
  • Honda
  • Kia
  • NIO
  • Nexdata

the data line, opened

One specimen. Five modalities. One gate it has to pass.

Every item is recorded where the product runs, reviewed by someone native to the language and qualified in the domain, and leaves with a manifest you can verify against the batch.

in your industry it lands as in the cabin, the ward, the counter, the line, the citizen's language and the course cabin-proven corpora and evaluation setsrights-clean clinical corpora with human reviewspecialist evaluation sets and document ground truthannotated inspection and QC data from the running sitemultilingual language data that serves every citizenlearning data reviewed by licensed educators and evaluated per subjectthe corpus, evaluation set and annotation your constraint demands

The corpus bench · one review gate, five modalities

speech · Healthcare

follow-up in six weeks, repeat the bloods before the visit

wakecommandoverlap

transcribed · diarised · PII tokens held

native reviewer · domain qualified accepted batch 0412 manifest sha-256 computing

Recorded where it runs

cabins, wards, lines and counters

Reviewed by people who speak it

native and domain-qualified review on every batch

A manifest at delivery

batch, rights and a hash for every item

AI data and evaluation services

the system line, opened

Cited, gated, chained. Edit one word in the source and the answer withdraws.

Every answer is a passage from an approved source. Everything beyond that goes to a person. Each log line carries the hash of the one before it.

in your industry it lands as in the cabin, the ward, the counter, the line, the citizen's language and the course in-cabin voice workflows the data servesworkflow assistants that respect the rights questionreviewable document workflows and controlled internal assistantsoperational workflows around inspection and QCcited citizen and caseworker assistantsassessment support that leaves the decision with the educatorthe assistant, agent or workflow your constraint allows

The controlled workflow · approved sources, a person in the gate, a chained log

approved sources Credit policy v4.2 · specimenApplication 0412 · specimen

A completed application, a registered-address document and an identity document. Security as specified in annex A.

Credit policy v4.2 · §3

the gate

decision log · hash-chained · each line carries the previous hash

  1. genesis · model 2.4.1 · policy v4.2 computing

Answers only from approved sources

cited inline, refused when the evidence is missing

Evaluated on your questions

the same turns, with hard variants, form the release gate

A chained record

every decision, model and policy version, on the log

AI Implementation services

What we produce, industry by industry.

01 · Automotive & Mobility

The cabin itself: noise, light, languages and real driver behaviour that rarely match a training set.

Voice that ships to a cabin is proven in one before release. We record, annotate and evaluate against your acceptance criteria, and build the in-cabin voice workflows the data serves.

in-cabin voice · real cabin noise

booth accepted · cabin false reject

record · sha-256 computing

System

In-cabin voice workflows built on the corpus, wake word to intent, inside the vehicle's HMI.

Data

Cabin-recorded corpora and evaluation sets, native speakers, seat and noise metadata on every turn.

delivered by AI Data & Evaluation · AI Implementation

  • Cerence AI
  • Hyundai
  • BYD
  • Honda
  • Kia
  • NIO
  • Nexdata

The same utterance, twice

click a source to mute it · drag to move it · the mic is at the dash

booth corpus · quiet room

“start route guidance”

start stop route root guidance …

accepted in evaluation

cabin at 110 km/h · snr -1 dB

“start route guidance”

start stop

false reject

run it on your own · moonshine tiny · in this tab

runs in this tab · weights load from ypai.ai on demand

evaluation record · sha-256 computing synthesised specimen

02 · Healthcare

The rights question: what documentation, speech and imagery a project may lawfully use.

Nothing enters a clinical corpus without the right to be there. We collect and review clinical speech and documentation under that question, and build the workflow assistants that respect it.

clinical documentation · lawful use only

consent on · in corpus · rights attached

record · sha-256 computing

System

Documentation and workflow assistants that respect the rights object on every turn.

Data

Rights-clean clinical speech and documentation, de-identified, reviewed by people qualified in the domain.

delivered by AI Data & Evaluation · AI Implementation

Nothing exists before the switch

mic off · capture path closed

run it on your own · electra small · in this tab

runs in this tab · weights load from ypai.ai on demand

one specimen line · your microphone or ours

follow-up in six weeks for Kari Nordmann, repeat the bloods before the visit on 4 March.

born with its rights object

purpose
model evaluation
basis
consent · art. 6(1)(a)
retention
90 days from capture
source
this session · specimen

into corpus · locked until review

03 · Financial Services

Risk and compliance cannot approve a workflow they cannot inspect.

A workflow risk can approve is one that keeps its record. We build document processing and controlled internal assistants that log input to output, evaluated against your policies and precision thresholds.

document processing · a reviewable record

address mismatch lit · routed to review

record · sha-256 computing

System

Document processing and controlled internal assistants that log input to output.

Data

Extraction ground truth, exception taxonomies and evaluation sets built from your own cases.

delivered by AI Implementation · AI Data & Evaluation

Input to output, on the record

application for credit facility · page 1 of 4 · specimen

The applicant requests a working-capital facility and confirms the information given below.

applicant A. Lovelace

registered address 14 Verkstedgata, 0552 Oslo

facility details · page 2 of 4

Purpose, security and repayment terms as specified in annex A.

requested facility NOK 240 000

identity document, copy · page 3 of 4

Issued 2021-05-02. Validity confirmed against the register.

address on identity document 7 Løkkeveien, 4008 Stavanger

declarations · page 4 of 4

The applicant declares the information complete and correct.

signed specimen signature

  • applicant A. Lovelace
  • registered address 14 Verkstedgata, 0552 Oslo differs from the identity document
  • requested facility NOK 240 000
  • identity document 7 Løkkeveien, 4008 Stavanger

decision log · append only

run it on your own · electra reader · in this tab

one question per field, answered as a span of the text · a rule compares the two addresses

runs in this tab · weights load from ypai.ai on demand

04 · Industrial & Energy

The site is already running. The data has to come from it, and the system has to fit what is there.

We capture and annotate inspection video and quality-control imagery on the working line, and build the operational workflows around the MES, the cameras and the shifts already in place.

site inspection video · systems already running

severity 2 · MES matched · operator hold

record · sha-256 computing

System

Operational workflows around inspection and QC, matched to the MES work order.

Data

Inspection video and QC imagery captured on the running line, annotated to the pixel.

delivered by AI Data & Evaluation · AI Implementation

One frame, classified against the line

inspection camera · line 4 · shift B · frame 0418

area
1 810 px²
size
3.4 mm
severity
2 · rework
pixel pitch
0.08 mm / px

MES work order WO 88213 · snapped · coordinates inside station 4

locked to the frame

drag the mask · the millimetres follow the geometry · release it inside the station to snap the order

run it on your own · slimsam · in this tab

mask confidence ·

a point prompt to a segmentation model · the mask it returns feeds the same millimetres, severity and work order

runs in this tab · weights load from ypai.ai on demand

inspection record · sha-256 computing

05 · Public Sector

An agency's assistant may only say what is approved, and must answer to every citizen.

We build knowledge systems and assistants that cite their approved sources and route the rest to a caseworker, and the multilingual data that lets them serve every citizen. Civilian use cases only.

citizen assistant · approved sources only

sealed · verified · out of scope routed

record · sha-256 computing

System

Citizen and caseworker assistants that cite approved sources and route exceptions to a person.

Data

Multilingual speech and text recorded by native speakers, evaluated per language before release.

delivered by AI Implementation · AI Data & Evaluation

One answer, sealed to its bytes

Housing benefit regulation 16:04 · after hours

Applications are assessed against the regulation in force on the date of receipt. The following provision applies to late submissions.

Late applications are accepted up to 14 days after the deadline where the delay is documented.

Applications received later than this are assessed only where statutory exceptions apply.

seal · sha-256(question ‖ passage bytes ‖ version ‖ approved time) computed at approval verified

journaled · 16:04 · after hours

verify this yourself

the seal is plain sha-256 · first 16 hex · run it in any terminal

edit one word in the passage and the seal breaks ·

run it on your own · minilm retrieval · electra reader · in this tab

§1 §2 §3 §4 §5 §6 §7 §8

dense retrieval over the approved passages, then an extractive reader on the best one · the answer is a passage from the source

runs in this tab · weights load from ypai.ai on demand

06 · Education

An assistant may support a grade, never issue one. Every point it awards traces to the learner's own words.

We build tutoring and assessment workflows that match an answer to the rubric and hand the decision to the educator, and the learning data they run on, reviewed by licensed educators and curriculum specialists, with the Article 10 evidence a high-risk classification asks for.

assessment against a rubric · the educator decides

two points matched · one open · the educator decides

record · sha-256 computing

System

Tutoring and assessment assistants that match a learner's answer to the rubric and leave the decision with the educator.

Data

Learning interactions, item pools and multilingual speech annotated by licensed educators and evaluated per subject.

delivered by AI Implementation · AI Data & Evaluation

The rubric · every point traced to the learner's own words, the educator decides

the question Explain why a control group is needed in an experiment.

the learner's answer · specimen

A control group gives you a baseline to compare the treated group against. Without it you cannot tell whether the change came from the treatment or from something else that happened during the study. The study also needs enough participants to be meaningful.

  1. 01 Isolates the effect of the treatment from other variables. matched · sentence 2 · 0.61
  2. 02 Provides a baseline to compare the treated group against. matched · sentence 1 · 0.86
  3. 03 Accounts for improvement caused by expectation alone, the placebo effect. open · 0.37

proposed 2 of 3 the educator decides

the educator may award or withhold any point

assessment record · sha-256 computing

run it on your own · minilm · in this tab

each rubric point is matched to the closest sentence by embedding similarity · a point below the threshold stays open · the educator confirms

runs in this tab · weights load from ypai.ai on demand

one team, two lines

When the failure sits in the data, the team that built the system fixes the data.

A corpus that never met the cabin. A document type missing from the ground truth. An evaluation set that does not represent production. One team owns the system and the data it runs on, and either line can be bought alone.

The AI Implementation service line Build the system. Assistants, agents, document AI and workflow automation, delivered into your operating environment. assistants and agentsdocument AIworkflow automationevaluation and monitoring The AI Data & Evaluation service line Build the data advantage. Speech, video, image, text and sensor data recorded where your product runs, reviewed and evaluated through its lifecycle. collectionannotationhuman reviewevaluation sets

the loop

  1. 01 Observe the failure
  2. 02 Isolate the gap
  3. 03 Produce the data
  4. 04 Improve the system
  5. 05 Re-run the evaluation
  6. 06 Release

who decides what

Run by YPAI Retained by you
QA sampling and review, run by YPAI. Acceptance criteria are yours.
Capture and labelling operations. Data, region and erasure stay yours.
Built into your operating environment. Configuration authority stays yours.
The pipeline, formats and records. Approval gates stay visible at every stage.

where it runs

Built and run from Norway, under the rules you already answer to.

YPAI is a Norwegian company operating under GDPR. The system runs in your environment, inside the access boundaries you already enforce. Every engagement runs on Controlled Delivery, with acceptance criteria, records and gates you can inspect.

Jurisdiction
Norway, under GDPR
Data residency
EEA by default, defined per project
Deployment
Private and EEA, in your environment
Processing terms
Standard DPA, subprocessors and transfers defined per project
Erasure
30-day end-of-contract erasure
EU AI Act
Article 10 documentation, produced as part of delivery

perspectives All perspectives →

The rules come first. Three we bring to every first meeting.

compliance · 7 March 2026 GDPR and AI: Enterprise compliance requirements → GDPR applies directly to AI training data collection, model outputs, and automated decisions. What enterprise compliance officers must address in 2026.
compliance · 16 December 2025 EU AI Act Article 10: Data Governance Checklist → infrastructure · 17 December 2025 CTO's Guide to Sovereign AI: Architecture & Costs →

workflow · constraint

Two sentences set the route.

The workflow you run, and the constraint it runs under. As you write, the route reads itself from your words. A named project lead confirms it and replies with feasibility.

The brief Two sentences are enough. The rest is scoped together.

YPAI will use the information provided to evaluate and respond to your scoping request. Privacy

Prefer to talk first?

first conversations

Five questions we hear before scoping.

Everything else is answered in scoping, with a named project lead.

Is my industry not shown here?

Every industry runs on the same two service lines and the same Controlled Delivery. The ones shown are where we can show the work in depth. Describe the workflow and the constraint, and a project lead sets the route.

Can our data stay in the EEA, or inside our own environment?

Yes. Processing, access, storage, review and delivery controls are designed around the data classification and your requirements, and defined before access begins. EEA-based, client-environment, private cloud, on-premises and hybrid configurations are assessed per use case.

Data residency in the EEA
Do we have to buy both service lines?

Either line stands alone. Where the work depends on both, YPAI scopes them as one engagement so the team that built the system also owns the data it runs on.

The two service lines
What does a pilot demonstrate?

Output quality, the proposed workflow, the review architecture, formats and the agreed acceptance criteria, against your requirements. Scope and commercial terms are agreed before the pilot starts, and production is a separate decision after review.

How a pilot runs
Who reads what I write in the form?

A named project lead. They read the workflow and the constraint, set the route to one line or both, and reply with feasibility and a first read on the constraint.

also on the same two lines

the route

You name the workflow and the constraint. We set the route, and prove it where it runs.

Scope an engagement Explore industry workflows

page record · sha-256 computing