AI-data services for vertical AI startups

Expert-graded AI data services.
Output, not surveillance.

Per-project SOWs from $30K. Senior raters with audited backgrounds. No screen capture, no biometric vaults, no Insightful-style surveillance.

Currently accepting projects in legal · medical · finance · sales verticals. Non-PHI, non-privileged.

The privacy promise

We grade output. Not contractors.

Maven Mark's quality assurance happens at the deliverable layer via peer review and rubric scoring. We do not surveil the people doing the work.

No screen capture

Deliverable-graded, not surveillance-graded. We never capture screens, keystrokes, or personal-app activity. Contractor privacy is policy, not preference.

Per-project PII compartmentalization

Your project, your data, your retention window. No centralized vault — segregated buckets per pilot, deletion windows defined in the SOW.

MFA-enforced + audited compliance

MFA on every contractor account from day one. Compliance posture via Vanta. We don't outsource to fraudulent vendors.

Service tiers

Three tiers. Match the rater to the rubric.

Blended hourly rates including platform fee. Final pricing locked in your SOW after a scoping call.

Vetted

$50/hr blended

Junior + mid raters on long-tail tasks. Throughput-optimized.

  • Pre-screened domain familiarity
  • Standard rubric grading
  • Anti-cheat scanning on every submission

Senior

Recommended

$120/hr blended

Domain experts on critical rubric work and SFT-grade rewrites.

  • 5+ years domain practice
  • Multi-rater quality with IRA reporting
  • Inline rubric authoring as part of work

Principal

$300/hr blended

Ex-AI-lab or ex-domain-leader for adjudication and corpus design.

  • Top-of-field credentials
  • Final adjudication on senior disagreements
  • Custom evaluation harness design

Most pilots blend two tiers. Pilot SOWs start at $30K; the typical first engagement lands between $50K and $120K. Monthly retainers from $25K + 35% platform markup.

How a pilot works

Four steps. First batch within two weeks.

01

Discovery call

30-minute conversation. We map your task family (preference pairs, rubric-graded eval, expert rewrite, executable benchmark), volume, vertical, and timing.

02

Scoped SOW

Within 5 business days you receive a fixed-bid SOW with deliverable count, contractor bench size, rubric draft, integration plan (S3 / webhook / REST / MCP), and pricing.

03

Bench assignment

We assign contractors from our vetted pool — Vetted, Senior, or Principal tier per the SOW. Bench-paid in advance so availability is real, not hypothetical.

04

Deliverable stream

Daily or weekly batches delivered via your preferred channel. Project status + IRA dashboard live throughout. Final corpus ships in your chosen format (HF DPO JSONL, OpenAI Evals YAML, Inspect AI, Argilla, SWE-bench, or custom).

What we deliver

Four task families. Every standard format.

No proprietary schema lock-in. We ship in whichever format your training pipeline already speaks.

RLHF / DPO

Preference pairs

Chosen / rejected response pairs for reward model training. Shipped in HuggingFace DPO JSONL standard — drop directly into TRL pipelines.

HuggingFace DPO JSONLConversational JSONL with roles

Capability evals

Rubric-graded evaluation

Multi-criterion scorecards on model outputs. Customer-authored or in-house authored rubrics. Met/Not-Met, 1-N rating, ranking, span highlighting.

Argilla recordsOpenAI Evals YAML+JSONLInspect AI Task module

SFT / Constitutional

Expert rewrites

Senior practitioners rewrite model outputs with rationale. Optionally net-new authored content for SFT corpora in specialized domains.

JSONL with original / edited / rationalePer-author trace metadata

SWE-style — v1

Executable benchmark tasks

SWE-bench-compatible task instances with golden patches, F2P / P2P tests, and Docker-runnable environments. Available in v1 for software-engineering customers.

SWE-bench JSONLArchipelago ZIPMCP server endpoint

Integration: S3 batch dump (default), signed webhook on milestone, REST API for pull (v1), or MCP server endpoint for agentic consumers (v1).

The question we expect to be asked

"If your competitor was breached and centralized contractor PII is in the news — what makes Maven Mark different?"

We grade deliverables, not contractors. We never adopted personal-device screenshot capture, biometric centralization, or W-9 vaulting as a product strategy.

We compartmentalize PII per-project. We enforce MFA from day one. We retain only what the SOW says, for as long as it says.

Privacy and security aren't a sales talking point we adopted in April. They are how the product is shaped.

Request a pilot

Tell us about your project.

One founder reads every request personally. Discovery call within one business day.

We never share your information. No marketing emails — we'll reach out once, personally.