Financial RLVR rubrics & traces

Helping AI take on financial services

Finance is the next frontier of AI value — and it is still too high-dimensional to grade itself. We build secure, audited pools of financial professionals who produce the human-verified rubrics, judgments and RLVR traces frontier labs and training-data platforms need.

Talk to us See the offering

TRACE-4471CPA · Rubric v3

Task

Restate FY24 EBITDA after reclassifying the sale-leaseback and capitalized cloud spend. Show the bridge.

Lease treatment matches ASC 842 3/3

Bridge reconciles to reported figure 2/2

Cloud spend capitalization unsupported 0/2

Assumptions stated and auditable 2/2

Verified score 7 / 9

Reviewer signed · chain of custody logged

$5B+

Frontier-lab spend on human data, growing ~50% YoY

24/7

SEA hours cover US night-time grading windows

450M

English speakers across Southeast Asia to recruit from

⅓–½

Fully loaded cost of equivalent US professionals

Why finance

The highest-value domain that AI still cannot grade on its own

Code and mathematics got verifiable rewards early because correctness is cheap to check.
Finance is the opposite: enormously valuable, heavily penalised when wrong, and still dependent on expert judgment to say whether an answer is right.

Legible Adjacent to mathematics — reasoning steps can be written down and scored Gradable

Valuable Sits directly on top of the white-collar work labs most want to automate High ARPU

High stakes Regulatory environments and penalties abound; errors are not cosmetic Regulated

Still high-dimensional Not RLVR-able today — it needs human grading and human-authored traces Human-in-loop

That combination is why finance is the low-hanging fruit — and why middlemen and labs will need a supply partner who has skilled contractors ready and available.

The landscape

Billions are already allocated, and demand keeps growing

Human labeling was the first wave. The next demand centre is generating RLVR and RLHF traces for further model learning, and it's driving billion dollar valuations in companies already.

Frontier-lab spend on human labeling Growing roughly 50% year over year $5B+

Anthropic, discussed spend in this space over the next year ~$1B

Google DeepMind, reported spend in this space $100M+

xAI, stated intent to scale investment in data and environments 10×

OpenAI raise to scale compute and data infrastructure February 2026 $110B

Our initial buyers are the middlemen

A VC-funded layer of 10–15 companies now sits between the labs and the expert supply: Fleet AI, Afterquery, Hud, Huzzle Labs, Vals AI, Idler, Aviro and others. Several have an explicit finance focus. All of them need professionals faster than they can recruit them.

Then the labs directly

Serving the middle market first gets us revenue, rep, and direct sight of the standards the frontier labs are actually enforcing.

Supply is the bottleneck

Money and demand are not the bottleneck. The real gap is vetted financial professionals who will do skilled rubric work, at a defensible cost, inside an environment a lab's security team will approve.

What we do

Human-verified financial grading, inside an audited architecture

We build pools of financial experts who output human-verified grading, rubrics and RLVR traces — and we wrap them in the secure, audited data pipeline.

Supply

Build the human teams

CPAs, auditors, equity and quantitative analysts, wealth managers, forensic accountants and CFOs — recruited, vetted and managed by a team that has built RLHF and annotation workforces before.

Custody

Secure and audited environments

Hardened data pipelines and infrastructure with logged chain of custody. Clients offload liability and receive standardised judgments, rubrics and RLVR traces in the formats labs and middlemen already use.

Compounding

Iterate toward the labs

Working in the frameworks the frontier labs want produces early revenue and a published benchmark — and positions us to serve those labs directly as volumes grow.

The SEA advantage

Affordable professionals, verified outputs, night-time coverage

Southeast Asia gives us a deep, English-speaking professional bench.

Fully loaded compensation vs. US equivalent ⅓ – ½

Regional population to recruit from ~700M

English speakers in region ~450M

Coverage of US night-time hours Full

Southeast Asian countries we recruit from

The team

Startups, finance, regulated tech, and annotation at scale

Four operators who have each already done one of the hard parts of this business somewhere else.

Cameron Olson

Chief Executive Officer

10 years founding and growing businesses, with 7 years in finance, technology, and strategy & data science leadership.

Akbar Hosseinkhani

Chief Technology Officer

Currently at Bloomberg. 15 years leading enterprise engineering teams through demanding architectural builds in highly regulated industries.

Bogdan Nedanov

Head of Product

13 years leading data science and data infrastructure teams at Stripe, Deloitte and Capital One.

Alanna Olicker

Chief Operating Officer

12 years running AI annotation and RLHF programmes for FAANG clients, leading projects of thousands of contractors.