THE PROBLEM

02

Someone has to answer: Can we responsibly let this AI do this job?

A company builds an AI that decides whether a payment should be paid, held, or sent to a person. It scores 98%. The one payment it should have held is inside that 98%.

WHAT THEY HAVE

A score

A labeling vendor's spreadsheet. A benchmark slide. A meeting where someone is asked to say yes.

WHAT THEY LACK

A defensible record

This version. The cases that matter. What changed. Who reviewed it. What evidence supports the decision.

WHO FEELS IT

The executive

The person who eventually has to answer: why did we allow this AI to do this job?

The problem is not that AI has no score. The problem is that the company cannot confidently defend putting it into the job.

EXPERTLUMA

02 / 12

THE SOLUTION

03

We run the production around the AI.

The customer brings the AI. Expertluma handles the controlled work required to produce reliable inputs, measure behavior, assemble evidence, and support the authorization decision.

01 · TRAIN AI

Prepare the material

Ingest data. Clean it. Detect sensitive data. Label examples. Review and accept the work. Preserve provenance. Release accepted training/evaluation material.

02 · VALIDATE AI

Prove the behavior

Test the exact AI version against a defined examination. Measure outcomes. Surface critical failures. Separate measurement from human review. Compare versions.

03 · CERTIFY AI

Support the decision

Bind evidence to this AI version and this job. Check policy, review, identity and conditions. The customer decides: authorized, not authorized, or not yet.

We do not build or train the customer's model. We operate the production around it.

EXPERTLUMA

03 / 12

THE PRODUCT

04

One AI. One job. One production record.

Here is what happens when a company brings Expertluma a real AI.

CUSTOMER AI

Clinical image model V2.4 Customer owns the model.

TRAIN AI

Prepare 50,000 images Label · inspect · accept

VALIDATE AI

200-case examination 197 correct · critical failures

CERTIFY AI

Evidence · policy · conditions Customer authorization

THE OUTPUT

An evidence-backed record of what was produced, what the AI demonstrated, what failed, and what decision applies to this exact version and job.

EXPERTLUMA

04 / 12

TRAIN AI

05

Prepare the material the AI needs.

Example: a healthcare company wants an AI to identify pneumonia from chest X-rays.

1

INGEST

Images, documents, audio, video or other source material

2

PROTECT

PHI/PII detection, quarantine and policy checks

3

LABEL

People produce the expected answer or classification

4

INSPECT

A different person reviews; quality and evidence gates apply

5

RELEASE

Accepted material is packaged with identity, provenance and checksums

CUSTOMER TRAINS THE AI

Expertluma prepares and controls the production material. It does not become the customer's model trainer.

EXPERTLUMA

05 / 12

VALIDATE AI

06

Measure what the exact AI actually does.

Same healthcare example. The customer submits AI version V2.4. Expertluma runs a sealed 200-case examination.

197

CORRECT

2

WRONG

1

CRITICAL UNRESOLVED

98.5% SCORE

STILL NOT READY

THE EXAM

Controlled

Cases are defined, locked and tied to the intended job.

THE REVIEW

Independent

A measurer cannot simply approve their own unresolved case.

THE RESULT

Evidence, not hype

A high average does not erase a critical unresolved failure.

Validation measures. It does not authorize.

EXPERTLUMA

06 / 12

CERTIFY AI

07

Can we responsibly authorize this AI for this job?

Certify does not create a new score. It binds the decision to the exact AI, intended use and evidence snapshot.

AI IDENTITY

Exact artifact fingerprint

VALIDATION

Completed measurement + evidence hash

CRITICAL SAFETY

No unresolved critical failure

HUMAN REVIEW

Required reviews complete

POLICY

Allowed for this use

DECISION

Authorized decision-maker approves

CUSTOMER DECIDES

AUTHORIZED · NOT AUTHORIZED · PENDING

A generic program approval cannot authorize a different AI version. A renamed version is not a new identity.

EXPERTLUMA

07 / 12

PROOF

08

Don't believe the diagram. Make me show you.

WHAT THE PRODUCT ALREADY DOES

• A person does the work. A different person accepts it.

• A serious failure can stop a high score.

• A renamed model is not a new AI.

• Validation does not authorize.

• Certify cannot approve.

• The answer can be not yet.

Internal demonstration. Not customer proof.

Built: yes. Customer proof: no. The round exists to change the second line.

EXPERTLUMA

08 / 12

WHY THIS CAN MATTER

09

The valuable thing is not another AI score.

It is the controlled production system connecting work, evidence, failures, versions and the authorization decision.

PRODUCTION

Work becomes evidence

Every accepted example can trace back through source, dataset, work item, worker, submission, inspection, acceptance and release.

VERSION CONTROL

Old evidence does not leak

Artifact identity is fingerprinted. A new AI version requires new evidence. A renamed version does not inherit the old answer.

DECISION INTEGRITY

No score becomes an approval

Validation, certification eligibility and executive authorization remain separate. The system can say NOT YET.

Train AI → Validate AI → Certify AI → Customer Authorization

The production record is the connective tissue. The customer's AI remains theirs.

EXPERTLUMA

09 / 12

THE BUSINESS

10

A company pays. Then its AI changes. Then it pays again.

The initial wedge is a bounded AI Production Program: a defined AI, a defined job, a defined evidence requirement and a customer authorization decision.

FIRST PRODUCTION PROGRAM

$25–75K

Deposit to start production. Balance when the contracted result is delivered.

FIRST PROJECT

Prove demand

A real company pays for production around a real AI.

NEXT VERSION

Create recurrence

Model, prompt, tools or retrieval change. Old evidence no longer covers the new version.

The business hypothesis is simple: companies pay for a defensible answer, then return when the AI changes.

EXPERTLUMA

10 / 12

THE ROUND

11

I built it. I cannot sell it from Nairobi alone.

FOUNDER

Vincent Rotich

Technical founder. Built the production system. The missing capability is commercial, not another year of feature development.

THE COMMERCIAL SEAT

US-based

Commercial co-founder or first commercial hire. The job: serious enterprise conversations, the first paid production program, then the second.

$2.5M PRE-SEED

SAFE · $10–15M POST-MONEY CAP · 18 MONTHS

Use of funds: US commercial seat · enterprise/security readiness · customer delivery · hosting · legal · runway.

We are raising before customers exist because the product is substantially built and the next risk is commercial.

EXPERTLUMA

11 / 12

THE BET

12

A company pays. Then its AI changes. Then it pays again.

One AI. One job. One evidence-backed decision.

We have built the production system. Now we need to find out whether companies will pay to use it.

SEE

A real AI enters production. The system can say NOT YET.

THEN JUDGE

Would an executive pay $25–75K for that answer?

THE ROUND

$2.5M · 18 months 2–4 paying companies

Vincent Rotich · Nairobi · mc.vickideas@gmail.com · expertluma.com

THEY KEEP THE AI. THEY DECIDE. WE RUN THE PRODUCTION.

EXPERTLUMA

12 / 12