01 / CREATE
Expert-authored training data
Original questions, reference answers, worked solutions, and structured examples built around your model's learning objectives. Specialist authors bring depth to reasoning-intensive tasks.
Expert-built data for frontier AI
Expert data.
Stronger models.
Expert-authored and synthetic data, mathematical formalization, and large-scale data processing for AI labs. Built by PhDs, Olympiad medalists, and domain specialists.
Specialized expertise. International collaboration. A focused free trial.
THE HUMAN LAYERDaEx / 01
Human expertise, built into every example.
01 Expert & synthetic data
02 Machine-checkable mathematics
03 Large-scale data processing
What we do
Better data starts with
people who know the field.
From original questions to generated examples, formal proofs, and large-scale preparation. Apply expert data and evaluation to frontier reasoning and robotics.
02 / GENERATE
Synthetic data
Expert-guided generation of problems, examples, and task variants. We design generation workflows around your target skills, then apply validation, filtering, and review before delivery.
Explore this service03 / FORMALIZE
Mathematical formalization
Turn mathematical statements and proofs into machine-checkable artifacts. Scope informal-to-formal translation, proof development, and verification against an agreed proof system and library version.
Explore this service04 / EVALUATE
Challenging evaluations
Specialist-designed tests that probe reasoning gaps, expose meaningful errors, and distinguish a plausible answer from a correct one. Each task has explicit evaluation criteria.
05 / REFINE
Human review & QA
Subject-matter review, annotation, and quality checks to resolve disagreements and improve consistency. Review assumptions, reference solutions, and labels with the right specialist.
06 / PROCESS
Big-data processing
Prepare large datasets through ingestion, cleaning, deduplication, transformation, and validation. We scope scalable processing and structured delivery around your data, infrastructure, and quality requirements.
Explore this serviceExpert-guided generation
Generate broadly.
Keep what holds up.
Synthetic data can be model-generated, programmatically generated, or a hybrid. We agree on source permissions, target skills, and acceptance criteria before generation begins.
01
Define the task
Choose the domain, difficulty, coverage, and output format.
02
Generate & filter
Create variants, remove duplicates, and run task-specific checks.
03
Review & package
Apply the agreed expert-review plan and deliver examples with metadata and quality notes.
Depth where it matters
Five ML specialties.
One expert network.
Match the work to the right specialist, rather than asking a generalist to cover every field.
SPECIALTY / 01
Classical & statistical
machine learning
Expertise in statistical reasoning, model selection, generalization, and evaluation. Tasks connect mathematical foundations with the choices practitioners make.
Original tasks, explained solutions, and specialist review.
SPECIALTY / 02
Deep learning
& optimization
Tasks spanning neural architectures, training dynamics, gradient-based optimization, and debugging. Specialists examine both the method and the reasoning behind it.
From conceptual understanding to technical failure analysis.
SPECIALTY / 03
NLP & large
language models
Language-model tasks designed around attention, adaptation, inference, evaluation, and reasoning. Review goes beyond fluent output to assess technical correctness.
Technical depth without confusing fluency with accuracy.
SPECIALTY / 04
Computer vision
& multimodal reasoning
Image-based tasks, visual interpretation, and model-evaluation problems. Work can combine figures, diagrams, code, and written reasoning in a single example.
Visual reasoning and reviewed inputs, including robotics applications.
SPECIALTY / 05
Reinforcement learning
& sequential decisions
Specialist tasks in reward design, value estimation, policy optimization, and sequential decision-making, with attention to assumptions and failure modes.
Sequential-decision tasks, including planning and robotics evaluation.
Frontier-model experience
Olympiad-level reasoning.
Research-level ambition.
Expert work for demanding mathematical, algorithmic, and scientific reasoning.
Our contributors bring experience supporting frontier-model training and evaluation programs aimed at gold-medal-level performance on IMO- and IOI-style challenges and other high-school STEM Olympiads. We connect that experience with expert-authored tasks, verified solutions, and rigorous review for advanced and research-level work.
Deep mathematical reasoning
IMO-style problems, rigorous proofs, counterexamples, and advanced mathematical tasks. Reference solutions make assumptions and logical steps explicit.
Algorithms that stand up to tests
IOI-style programming challenges with algorithmic explanations, complexity analysis, reference implementations, and tests matched to the task.
From principles to solutions
Physics, chemistry, and other STEM reasoning tasks scoped to the right specialists, with attention to assumptions, units, derivations, and failure modes.
Contributor experience describes individual project work, not medals awarded to DaEx. Model targets, test conditions, and success criteria are agreed per engagement; benchmark outcomes are not guaranteed.
Discuss a reasoning objectiveMathematics & formal verification
Mathematics, made
machine-checkable.
We translate informal mathematics into precise formal statements and develop proofs that can be checked in the agreed environment. Projects can include theorem-and-proof pairs, formalization datasets, proof repair, and evaluation tasks.
We check both the proof and whether the formal statement captures the intended mathematics. Definitions, assumptions, dependencies, and acceptance criteria are part of the specification.
PROOF SYSTEMS
The proof system, version, libraries, and specialist availability are agreed for each project.
A REVIEWABLE FORMALIZATION PACKAGE
01
Statements & proof files
Formal definitions and proofs, paired with the original mathematical intent.
02
Validation records
Proof-check results, dependency information, and an explicit record of assumptions.
03
Reproducible environment
Versioned source files and instructions to re-run the agreed checks.
Artifacts labeled as verified must meet the agreed checker requirements, with no unresolved proof holes or undisclosed assumptions.
Discuss a formalization projectData engineering at scale
Large datasets.
Clear quality controls.
We process large collections of text, code, images, documents, and structured records into consistent, usable data for AI workflows.
We agree on volume, compute environment, processing rules, and delivery targets before work begins. A small representative batch establishes the checks before the workload expands.
Discuss your processing requirements01
Ingest & structure
Import approved sources, parse files, standardize schemas, and preserve source metadata.
02
Clean & validate
Normalize records, remove duplicates, flag malformed content, and apply agreed quality checks.
03
Process at scale
Use partitioned workloads and distributed batch processing matched to the project infrastructure and throughput targets.
04
Deliver & document
Export formats such as JSONL, Parquet, or CSV, with dataset versions, manifests, and quality summaries.
Data access, retention, security requirements, and infrastructure costs are agreed before processing. The public inquiry form is for project summaries, not dataset uploads.
Robotics & embodied AI
Data for perception,
planning, and action.
Extend expert review and scalable data workflows to models that reason about and interact with the physical world.
PERCEIVE
Visual & spatial understanding
Annotate client-provided images and video for objects, spatial relationships, interaction opportunities, and grounded scene descriptions.
Possible outputs: reviewed labels, captions, spatial questions, and evaluation rubrics.
PLAN & ACT
Demonstrations & action sequences
Curate recorded demonstrations, break activities into task stages, align instructions with observed actions, and review trajectories and failure cases.
Possible outputs: annotated sequences, step-level labels, and instruction-action examples.
EVALUATE
Simulation data & task evaluation
Scope synthetic examples and simulation-data curation with your engineering team. Build review criteria for navigation, manipulation, and task completion.
Possible outputs: scenario specifications, curated batches, success criteria, and error analyses.
We agree on modalities, source permissions, annotation standards, infrastructure, and validation requirements before starting. Physical data collection, robot hardware integration, and real-world deployment require separate scoping.
Discuss a robotics projectThe people behind the data
Gold medalists. PhDs.
Research meets engineering.
Our team includes IMO and IOI gold medalists pursuing degrees in computer science and AI, graduates with double majors in mathematics and computer science, and mathematics PhDs and postdoctoral researchers interested in AI and development.
Academic backgrounds across our network include MIT, Caltech, UC Berkeley, Georgia Tech, and the University of Miami. We match contributors to the subject, toolchain, and review needs of each project.
Our network includes experts with experience on Mercor and Handshake projects. Individual contributor experience does not imply a partnership with, or endorsement by, those companies.
Frontier-model reasoning experience
Experience contributing to training and evaluation programs targeting gold-medal-level Olympiad reasoning, with work spanning mathematical problem solving, algorithms, and research-level challenges.
Technical review experience
Subject-matter expertise supported by annotation, quality assurance, and research-paper review experience across our network.
Mathematics & software engineering
Specialist teams for original reasoning tasks, synthetic-data workflows, formal proof development, and data-processing pipelines.
TEAM EXPERIENCE
Our team
Mathematics, AI & software
DaEx LLC
Current Meta work.
Previous Google experience.
CURRENTMeta AI projectsPREVIOUSGoogle projects
Our team's experience includes current contractor work on Meta AI projects and previous work on Google projects, spanning mathematical reasoning, coding evaluation, and expert review.
These are individual credentials from contractor project work. DaEx is an independent company; the references do not imply endorsement by Meta or Google, or a client relationship between those companies and DaEx.
ACADEMIC BACKGROUNDS ACROSS OUR NETWORK
MITCaltechUC BerkeleyGeorgia TechUniversity of Miami
Educational backgrounds describe individual contributors, not university partnerships or endorsements.
INTERNATIONAL REPRESENTATIVES
NORTH AMERICAUnited States
EAST ASIAChina
MIDDLE EASTSaudi Arabia
EUROPEGermany
Our approach
From research objective
to reviewable deliverables.
A clear scope and a shared definition of quality come before volume.
01 / ALIGN
Define the target
Agree on domains, difficulty, source permissions, data formats, processing scale, and what a correct deliverable looks like.
02 / CREATE
Choose the workflow
Match expert authoring, synthetic generation, formalization, data processing, or robotics-data review to the right specialists and tools.
03 / REVIEW
Validate the work
Apply task-specific review, automated checks, and proof verification where relevant. Document assumptions and resolve disagreements.
04 / DELIVER
Refine, then scale
Review a representative batch together, record feedback and acceptance criteria, then expand the agreed workload.
See the work first
One week.
Real expert work.
No trial fee.
Start with a focused sample engagement. Evaluate the depth of our expertise and the quality of our deliverables before discussing a larger project.
A practical way to assess the fit
A mutually agreed domain and sample scope
An agreed authoring, synthetic-data, formalization, or processing sample
A review of quality, feedback, and next steps
The one-week trial covers an agreed sample scope and depends on expert availability. Any work after the trial is separately agreed and priced. No performance outcome is guaranteed.
Before we begin
A few useful
answers.
Can your data services support robotics?
Yes. We can scope perception annotation, spatial reasoning tasks, recorded-demonstration curation, action-sequence review, and simulation-data evaluation. The work starts with your task, available inputs, and validation requirements. Hardware integration and real-world data collection require a separate agreement.
What kinds of data can we work on?
We can scope expert-authored questions, synthetic examples, reference solutions, annotations, scoring rubrics, model evaluations, formal proofs, and large-scale data processing. Inputs and outputs may include text, code, images, diagrams, and files.
How is synthetic data checked?
We define acceptance criteria before generation. Checks may include schema validation, deduplication, executable tests, proof checking where applicable, and an agreed level of expert review. Model-generated content is not treated as correct solely because it was generated.
Which formalization systems can we discuss?
Projects can be scoped in Lean 4, Isabelle/HOL, Rocq (Coq), Agda, or HOL Light. We agree on the proof system, versions, libraries, specialist availability, and proof-checking criteria before starting.
Can you work with large datasets?
Yes. We can scope ingestion, cleaning, deduplication, transformation, validation, and scalable batch processing. Volume, throughput, infrastructure costs, access controls, and delivery requirements are established for each engagement.
Do I need a project in all five ML domains?
No. Start with one specialty or a focused problem. We match the work to relevant experts rather than requiring a broad engagement.
Can we discuss confidential requirements?
Begin with a non-confidential overview. Agree on confidentiality, source permissions, data handling, and delivery terms before sharing proprietary datasets or unpublished research. Do not send sensitive files through this inquiry page.
Do you guarantee a benchmark score?
No. We offer expert work and agreed deliverables, not a guaranteed benchmark result. Any claim about model improvement should be supported by an agreed evaluation.
How do international engagements work?
Our representatives support international conversations and project coordination. We agree on the scope, contracting arrangements, permitted access, and delivery requirements before the work begins.
Let's build something rigorous
What should your
model do better?
Tell us your target domain and whether you need original data, synthetic examples, formal proofs, frontier-model evaluations, robotics data, or large-scale processing. Start with a focused free trial or discuss a larger engagement.