Step 01 / 05
Scope thework.
We agree the task, output format, acceptance checks and delivery schedule with you, in writing, before anyone starts.
Egocentric dataFirst-person video of real human tasks →
Signal you canverify
Egocentric video from the human point of view, plus benchmarks, training data and agent environments, built by vetted engineers, researchers and domain experts and reviewed before they reach your pipeline.
Egocentric capture · Expert annotation · Three review gates on every taskTASK · LT-EVAL-0417live review
instruction.md
Fix the failing build in auth/session.py
environment/Dockerfile · tests/verifier.py
01Automated screenPASS
02Rubric score · 5 dimensionsPASS
03Senior reviewer sign-offPASS
Delivered with review notes
Why Latryn
Most AI data gets checked after it ships. Ours is checked before it leaves, by people who know the domain, against rules you agreed to.
03 / How we work
How it works,
How it works,
in five steps.
Step 02 / 05
Calibrate on apilot.
A small pilot batch runs first so reviewers and your team agree on the edge cases and the quality bar.
Step 03 / 05
Build withmatched experts.
Work is routed to the specialists whose experience fits the domain, with independent review arranged from the start.
Step 04 / 05
Screen, score,sign off.
Every task passes the three gates. Issues are recorded, returned to the right person and checked again against the original requirements.
Step 05 / 05
Deliver, witha named contact.
You receive the dataset with review notes, delivery records and one person to talk to, not just a file.
POV / Egocentric data
Data from the
Data from the
human point of view
We record people doing real tasks through head-mounted cameras, then annotate what matters: hands, objects, actions and intent. Built for robotics, embodied AI and video models.
REC 00:00:12POV · head-mounted
mug · targetkettleright hand · approaching
reachgrasppourplace
“C reaches for the mug with the right hand.”
SCREEN PASSSCORE PASSSIGN-OFF PASS
02 / What we build
Five practices.
Five practices.
One quality bar
Pick one practice or run them together. Every task goes through the same three gates.
01 / Eval
Latryn Eval
Benchmarks and model evaluations that test the work a model actually has to do, with scoring criteria, reviewed responses and documented failure modes.
Scope an eval02 / Train
Latryn Train
SFT pairs, chain of thought, preference rankings and reward-model data from domain specialists.
Scope training data03 / Env
Latryn Env
Reproducible Docker and terminal environments for agent tasks, with verifiers that judge the result.
See sample tasks04 / Review
Latryn Review
Independent checks on instructions, code and results, plus red teaming that finds where models break. Every verdict comes with its evidence.
Talk to our team05 / Capture
Latryn Capture
Egocentric video from the human point of view. Contributors wear head-mounted cameras while they do real tasks, then trained reviewers segment and label every clip for robotics and embodied AI.
Scope a capture program
01 / Who we work with
One expert network.
One expert network.
Three ways in
For AI labs
Benchmark-grade evaluation and training data.
Harder benchmarks, contamination-resistant tasks and research-grade post-training data and egocentric video for robotics and embodied models.
Talk to a researcherOriginal tasks written to resist training-set leakage
First-person video annotated for hands, objects and actions
Rubrics and reviewer pools calibrated per program
Task-level status and domain coverage while work runs
For AI platforms
Domain-specific data at scale.
Human-evaluated datasets built around the capabilities your product needs and the post-training work behind them.
Talk to our teamData shaped to your product's real use cases
Works inside your task format, tools and guidelines
Continuous throughput once reviewers are calibrated
For enterprise AI teams
Reliable data for the systems you run.
Vetted, human-evaluated data pipelines for the AI workflows your teams depend on every day.
Talk to an expertScope, access and acceptance agreed before work starts
Delivery records showing what changed and why
A named point of contact for every program
04 / Quality
Calibrated signal, not crowd volume
Every task passes three gates before it enters your dataset. Work that misses the bar goes back, not out.
01PASS
Automated screen
Each submission is checked against defined criteria before it enters the pipeline. Failures are flagged immediately.
02PASS
Structured scoring
Tasks are scored across several quality dimensions. Every pass, flag or reject is backed by evidence.
03PASS
Human sign-off
Edge cases go to a senior reviewer. Nothing borderline ships without a person making the final call.
| Aspect | Typical crowd vendor | |
|---|---|---|
| Who does the work | Open crowd, mixed experience | Vetted specialists matched to the domain |
| Quality control | Spot checks after the fact | Three gates on every task |
| What you see | A finished file at the end | Task-level status, notes and records |
| Custom programs | Fixed templates | Built around your rubric and quality bar |
| When something fails | You find it and re-clean | Returned, fixed and re-checked first |
05 / Task library
See the work,
See the work,
not just the file
Read the brief, inspect the environment and run the verifier yourself.
Open the task libraryinstruction.mdbrief and expected output
task.tomllanguage · domain · scope
environment/Dockerfilereproducible env
fixtures/inputs/cases to handle
tests/verifier.pyacceptance checks
review/notes.mdreviewer sign-off
$ run verifier → 14/14 checks passed
06 / The Lattice
Work on the problems
Work on the problems
you already know
Our vetted network of engineers, researchers and specialists. Everyone passes a two-part timed assessment: an auto-scored technical section and a scenario section reviewed by our team.
Join the Lattice
Software EngineerAuthor hard terminal tasks with reference solutions and verifiers.
ML & Evaluation SpecialistDesign evaluation cases and score outputs against a rubric.
Domain ExpertReview tasks where correctness needs specialist knowledge.
Applications open
Apply to the Lattice
Five minutes: your background, the work you want to do and your resume. You get a confirmation email straight away.
07 / From the blog
Research-backed notes
All articles
Research-backed notes
on AI data
Egocentric data
Egocentric data collection for robotics and embodied AI
What Ego4D, EPIC-KITCHENS-100 and Ego-Exo4D teach about capturing and labelling first-person video.
EvaluationWhat makes an AI benchmark trustworthy
Contamination, unfair tests and vague task statements quietly break evaluations. How to build benchmarks you can believe.
Training dataRLHF and preference data, explained for teams that collect it
What to collect, who should collect it, and how DPO changes the pipeline.
08 / Before we start
Questions,
Questions,
answered plainly
Pick a question. Something else in mind? Write to our team.
Latryn teamusually replies within one working day
How do you make sure the data is good?
Every submission is screened automatically, scored against defined criteria across several dimensions, and reviewed by a senior reviewer whenever anything is borderline. Work that fails is returned with notes, not delivered.
Work with Latryn
Tell us the benchmark. We'll match the reviewers
Share the scope, expected volume and timeline. Setup usually takes a few days, and after that throughput is continuous. No commitment at the scoping stage.