Human infrastructure for AI agents & robots

The human layer for agents and robots.

Egocentric capture, expert annotation, real-time judgment, and teleoperation — one partner for the human side of your AI stack. When your agent needs a judgment call or your robot needs a hand, 1,200+ managed professionals answer. Not gig workers.

Live across our centers
Egocentric capture100s of hrs / day, 4K
Annotation throughput500M+ delivered
HITL responsesub-minute
Teleop benchesstaffed 24/7
Workforcemanaged, in-office
100s of hrs
4K egocentric video captured daily — and growing
500M+
Annotations delivered via IndiVillage
1,200+
Managed professionals — salaried, trained, in-office
99%+
QA accuracy · 98% client retention
The Stack

One partner. The whole human stack.

Everything your agents and robots need humans for — from pretraining data to live judgment to deployment fallback. Four products, one workforce, one API.

The Flywheel

Every intervention is a labeled demonstration.

Your hardest moments become your model's best data. A robot stalls — or an agent hits a question it shouldn't guess at. Our people resolve it live. The resolution comes back annotated and training-ready. Next quarter's model escalates less. That loop is the product.

1
Capture
Egocentric human video at scale seeds the model.
2
Train
Annotated, taxonomy-aligned data goes into pretraining.
3
Deploy
Your robots and agents ship before the model is perfect — safely.
4
Escalate
Edge case hit: the robot or agent requests a human via API.
5
Intervene
Judgment call or full teleop — resolved in seconds.
6
Annotate
The intervention returns as a labeled demonstration.
↺  back into training — fewer escalations every cycle
Seconds
from escalation to a trained human on task
100%
of interventions returned as training-ready data
One vendor
from pretraining data to deployment fallback
01 — Capture

Hundreds of hours of 4K egocentric video. Every day.

Robots learn manipulation from human hands. We run managed capture centers where trained collectors — on staff, not crowdsourced — record first-person video of real tasks: cooking, cleaning, folding, assembly, picking, repair.

You direct the taxonomy. We deliver the hours — raw, curated, or fully annotation-ready through our own labeling teams.

  • Already running. Hundreds of hours of 4K capture per day, scaling weekly.
  • Directable. Scripted task programs against your taxonomy, or naturalistic full-day capture.
  • Clean provenance. Consent-first collection, full metadata, PII scrubbing.
  • One pipeline. Capture and annotation under one roof — no vendor hand-off.
Request a Sample Dataset
Capture Spec
RigHead-mounted 4K @ 30/60 fps
StreamsVideo · audio · IMU (gaze optional)
CoverageScripted tasks or naturalistic days
Throughput100s of hours / day, growing
ProvenanceConsent-first · metadata · PII-scrubbed
DeliveryRaw · curated · annotation-ready
Custom programs: need a specific environment, tool set, or demographic mix? We stand up dedicated capture programs against your spec.
02 — Annotate

Annotation that's already proven at scale.

We don't claim a labeling capability — we point at one. Our annotation operation is run with our sister company, IndiVillage.

Powered by IndiVillage

IndiVillage has delivered 500M+ annotations for global AI teams from managed delivery centers — salaried professionals working in offices, trained per project, retained for years. Its impact-sourcing model builds tech careers in communities that rarely get them: quality your ML team can verify, and a workforce story your stakeholders can be proud of.

Egocentric video labeling 3D pose & hand tracking Object & affordance tagging RLHF preference ranking Agent trace evaluation Multi-turn conversation rating Content moderation Document extraction Custom taxonomy & ontology design
QA built in: dual-pass review, gold-set sampling, and calibration sessions hold delivered accuracy above 99%. Pilots start from as few as 10 hours.
03 — Judge

An API for human judgment.

Five primitives your agents and robots call when they need a human. One REST API or MCP server. Sub-minute responses with full audit trails.

Classify

Binary or multi-class categorization by trained humans.

Judge

Pairwise comparison, RLHF ranking, subjective evaluation.

Extract

Structured data from documents, images, or video.

Escalate

Route high-risk or low-confidence cases to specialists.

Resolve

End-to-end human resolution with rationale and audit trail.

Relay: complex questions, priced binaries.

Relay is the intelligence layer between your agent and our humans. It decomposes any question into binary decisions, routes each to the right tier, runs them in parallel, and reassembles the answer — with the full trace returned for your learning loop.

  • Negotiate. Already a binary? It routes directly. If not, Relay proposes a decomposition.
  • Decompose. Independent binaries run in parallel across workers.
  • Reassemble. Answers compose into a final response with reasoning.
Basic$0.50 / call Complex$1.00 / call Expert$2.50 / call
Your robot asks
"Should I replace this boiler or repair it?"
Relay decomposes
Is the unit more than 15 years old?Basic $0.50
Visible corrosion on the heat exchanger?Expert $2.50
Flue gas reading within safe limits?Expert $2.50
Relay reassembles
"Replace. Corrosion present, flue reading borderline, unit is 18 years old."
3 binaries · 2 in parallel · 47 seconds · $5.50 total
from humanrelay import HumanRelay hr = HumanRelay(api_key="hr_live_...") result = hr.judge( content={"a": response_a, "b": response_b}, rubric="Which response is more helpful?", tier="expert", ) print(result.verdict) # "Response A" print(result.rationale) # "Response A provides..."
curl https://api.humanrelay.com/v1/judge \ -H "Authorization: Bearer hr_live_..." \ -H "Content-Type: application/json" \ -d '{ "content": { "a": "...", "b": "..." }, "rubric": "Which response is more helpful?", "tier": "expert" }'
// claude_desktop_config.json / .mcp.json { "mcpServers": { "humanrelay": { "command": "npx", "args": ["-y", "@humanrelay/mcp"], "env": { "HUMANRELAY_API_KEY": "hr_live_..." } } } }
04 — Operate

A human hand on the wheel, in seconds.

Deployed fleets meet the long tail: a jammed gripper, an ambiguous object, a customer at the door. Operate puts trained teleoperators behind your robots around the clock.

The robot requests help. We confirm safety. An operator takes over. Control hands back. You get the resolution — and the recording, annotated, as a demonstration for your next training run.

  • Staffed benches, 24/7. Operators on shift in our centers — not on-call gig workers.
  • Latency-tiered routing. Standby SLAs matched to task risk.
  • Safety gate. A $0.50 confirmation binary before any consequential physical action.
  • Data exhaust included. Full intervention trace plus annotated demonstration, returned via API.
Discuss a Teleop Pilot
Intervention Lifecycle
1 · EscalationRobot calls the API with context
2 · Safety checkConfirmation binary, sub-minute
3 · SessionOperator takes control (low-latency)
4 · HandbackRobot resumes autonomous operation
5 · ReturnTrace + annotated demonstration
No fleet yet? We also run scripted teleoperation programs purely for data collection — on your rigs or ours.
The Workforce

Managed teams. Not gig workers.

Gig platforms churn anonymous workers through your data. We put named, trained, salaried professionals on your project — in offices we run — and keep them there.

Quality

Same people on your project every day. Project-specific training, calibration sessions, dual-pass QA. That's how you hold 99%+ accuracy and 98% client retention — numbers a revolving crowd can't reach.

Security

Access-controlled floors, NDAs, managed devices, SOC 2 compliance. Your pretraining data and customer footage never touch an anonymous crowd — every person who sees it is accountable by name.

Impact

Through IndiVillage's impact-sourcing model, this work builds careers in communities that rarely get them. Procurement will like the security posture. Your board will like the story.

Who it's for

Built for the people building agents and robots.

For agent builders

Judgment on tap for production agents.

Five HITL primitives behind one API or MCP server — moderation calls, RLHF, evals, escalation. Sub-minute human answers for the moments your agent shouldn't guess. From $0.50 a call.

Start a pilot →
For humanoid companies

Deploy before the model is perfect.

Pretraining ego-video at scale, evaluation pipelines, and a 24/7 teleop fallback behind every unit in the field — so the fleet ships now and improves monthly.

Start a pilot →
For frontier labs

Data you can put in the model card.

Diverse, consented, provenance-clean human data for VLA and world models. RLHF, rubric evals, and red-teaming by trained specialists — not a marketplace.

Request sample data →
For investors

The bottleneck is the business.

Embodied AI's constraint isn't compute — it's human data. Ask for the memo: traction, the IndiVillage moat, and the intervention flywheel.

Request the memo →
Pricing

Four products. Simple meters.

Pay for delivered hours, labeled assets, resolved calls, or completed interventions. No platform fees, no seat licenses.

Capture
per data-hour

Program-based egocentric capture, priced per delivered hour by spec.

  • Pilot batches to standing daily capture
  • Raw, curated, or annotation-ready
  • Custom environments & taxonomies
Annotate
per asset / per hour

Labeling and evaluation with QA included, via IndiVillage delivery centers.

  • Pilots from as few as 10 hours
  • Volume pricing at scale
  • Dedicated teams for standing work
Judge
$0.50 – $2.50 per call

Human judgment by API: Basic $0.50 · Complex $1.00 · Expert $2.50.

  • Sub-minute response targets
  • Relay decomposition for complex asks
  • Full audit trail on every call
Operate
standby + per intervention

Teleoperation SLAs tiered by latency and risk, plus per-intervention pricing.

  • 24/7 staffed coverage
  • Safety-gated control sessions
  • Annotated demonstrations included

Volume discounts at scale · SLA guarantees available · hello@humanrelay.com for a quote

FAQ

Questions, answered straight.

One partner for the human side of AI — software agents and embodied AI alike: egocentric data capture, annotation and evaluation (run with our sister company IndiVillage), a real-time human-judgment API your agents call when they shouldn't guess, and teleoperation for deployed robot fleets. One workforce of 1,200+ managed professionals powers all four.
Three ways. First, workforce: our people are salaried, trained, office-based teams — not an anonymous crowd — which is what holds 99%+ accuracy and lets us pass security review. Second, real time: sub-minute judgment calls and live teleoperation, not just batch labeling queues. Third, the full lifecycle: capture, annotation, judgment, and teleop under one roof, so every deployment escalation flows back into your training data without a vendor hand-off.
Head-mounted 4K at 30/60 fps with synced audio and IMU (gaze optional), covering real manipulation tasks: cooking, cleaning, folding, assembly, picking, repair. We're capturing hundreds of hours daily and scaling weekly. You can buy from standing programs or direct your own: your taxonomy, your task scripts, your environment requirements. All capture is consent-first with full provenance metadata and PII scrubbing.
IndiVillage is our sister company and the engine behind our annotation track record — 500M+ annotations delivered to global AI teams from managed delivery centers. It pioneered an impact-sourcing model: salaried professionals in offices, building tech careers in communities that rarely get them. HumanRelay productizes that workforce for the era of agents and robots.
Judge calls target sub-minute resolution, tiered by complexity. Teleoperation runs on standby SLAs: for fleets under coverage, a trained operator is on task in seconds. Annotation and capture programs run on agreed delivery schedules, from 48-hour pilots to standing daily throughput.
Your robot calls the escalation API with context (sensor frames, task state, requested control scope). A safety-confirmation binary gates the session. An operator connects over a low-latency channel, completes or unblocks the task, and hands control back. The full trace — video, control inputs, rationale — is returned via API as an annotated demonstration you can fold into training.
All work happens in access-controlled offices on managed devices, under NDA, with SOC 2 compliance. Capture is consent-first with provenance metadata on every clip and PII scrubbing in the pipeline. Because our workforce is employed rather than crowdsourced, every person who touches your data is identifiable and accountable.
Scoped and fast: Judge pilots return results in 48 hours; annotation pilots start from as few as 10 hours of work; capture pilots deliver a sample batch against your taxonomy; teleop pilots begin with a scoped coverage window. No commitment beyond the pilot. Email hello@humanrelay.com and we'll scope it this week.

Put 1,200 managed professionals behind your agents and robots.

Capture, annotation, judgment, teleoperation — scoped as a pilot this week.

No commitment required · Pilot results in 48 hours · SOC 2 compliant

Or email us directly: hello@humanrelay.com