/labs

Human intelligence infrastructure for frontier AI

2M+ verified professionals, against real credentials not self-report.
2.0M+
verified professionals
500k+
advanced degree holders
100+
domain experts
<24h
to a staffed expert panel
§ 1.0 / Capabilities

One expert layer for frontier models.

POSTlabs.rlhf·p50: 4h · p95: 22h

Reinforcement learning from human feedback

Preference ranking, supervised fine-tuning, and reward modeling pipelines sourced from credential-verified experts.

inputs
  • · model_outputs[]
  • · rubric
  • · expert_pool
outputs
  • · preference_pairs
  • · reward_signals
  • · rationales
§ 2.0 / Network

Distribution by professional domain.

#
domain
share
n
01
engineering
37.5%
750,000
02
business
17.6%
352,000
03
sciences
9.3%
186,000
04
finance
8.8%
176,000
05
medical
7.0%
140,000
06
creative
6.2%
124,000
07
legal
5.9%
118,000
08
other
7.7%
154,000
Σ
total
2,000,000
§ 2.5 / Early access

Early access.

Domains we are actively scaling. Request early access to scope an engagement.

rlhf.medicalearly access

Medical reasoning

Clinical scenario authoring and factuality evals from board-certified physicians.

MD/DO · 12 specialties · IRB process & HIPAA handling
request early access
rlhf.tool_useearly access

Tool use & agentic evals

Long-horizon agent trajectories, tool-call grading, and environment design for coding and workflow agents.

senior eng · multi-turn · trajectory-level
request early access
§ 3.0 / Sample data

Pick a domain — get 500 expert-authored examples.

labs.aiapply.dev— /samples
aiapply@labs ~$ curl -O "labs.aiapply.dev/samples/rlhf.legal.jsonl"
What are you interested in?
↑↓ navigate1–6 select↵ confirm
§ 7.0 / Pipeline

Expert → dataset → model improvement.

#
stage
operation
01
source_experts
Query 2M+ professional pool by domain, credentials, availability.
02
qualification
Skill assessments, calibration tasks, identity verification.
03
task_assignment
Matching engine routes by expertise, language, latency target.
04
quality_review
Multi-rater overlap, inter-rater κ, expert adjudication.
05
dataset_generation
Provenance-tagged JSONL with rater metadata and rationales.
06
model_improvement
Continuous evaluation loop wired to your training infra.

Request research access.