FAQ
Questions buyers ask us.
Short answers for AI teams evaluating Synthium. Where something is agreed per engagement rather than standard, the answer says so. Anything not covered here: talk to us.
29 questions
Working with Synthium
What does Synthium do?
Synthium supplies expert-led data work, reinforcement-learning environments, agent evaluation and authorized adversarial testing to teams building frontier AI.
These are four offerings (Data, Train, Agents and Red Team) that share one foundation: qualified experts, data operations and independent evaluation. Each is a way to work with us in its own right.
Which expert fields can you staff?
We recruit across software engineering, mathematics, computer science, finance, law, healthcare and medicine, science, languages and research.
Whether we can staff a specific profile at your quality bar is confirmed during scoping. It is not promised in advance.
Where are your experts based?
Our contributor network is built from India's deep technical talent pools, and we work remote-first with global delivery.
Location is how we reach depth in a field. What we provide is calibrated expert judgment, not headcount. If your project needs a particular language, region or credential, say so at scoping.
Can the work be adapted to our pipeline?
Largely, yes. Deliverable formats, rubrics, review chains and tooling are agreed at scoping so the output fits your pipeline.
What does not change is that experts are qualified and calibrated before they touch production work, and that output is independently reviewed.
Data and data partnerships
What does Synthium Data cover?
Synthium Data covers two kinds of work: assessing eligible company-owned data for acquisition or licensing, and turning data or domain tasks into expert-reviewed datasets through curation, annotation, quality assurance and adjudication.
If I submit my data, will Synthium acquire or license it?
Not necessarily. Submitting starts a review; most opportunities will not proceed, and a reply is not an offer.
Rights, provenance and quality come before any commercial conversation. Terms are not described before an agreement exists.
Should I send a sample or the data itself?
No. Describe the data in words. If we proceed, we agree a secure transfer route first.
Please do not send data, samples, credentials or personal records through a form or by email.
What do you check before considering a data opportunity?
We look at relevance, uniqueness, provenance, clarity of rights, quality, coverage, and whether the data may lawfully be used for AI training or evaluation as intended.
If you are not sure you have the right to license the data, say so on the form. We treat that as a question for the rights assessment and do not assume the answer.
RL environments
What is an RL environment, as Synthium builds it?
It is a runnable task with a defined starting state, a set of tools, explicit completion criteria and a verifier that is independent of the agent.
The agent takes steps and sees what they do. The verifier checks the resulting state and behaviour, not the wording of an answer.
Do you run the training, or do we?
You do. Training runs in your own stack. Synthium supplies the environment, tasks and verification.
The exact contents of a delivery, such as the environment specification, tasks, execution interface, verifier and feedback specification, are agreed for each engagement.
Do I need both environments and post-training datasets?
No. They are different assets, produced differently, and you can engage for either one.
They combine well. Expert demonstrations and critiques can shape the tasks an agent is trained on, and a verifier's failures can point to the examples worth writing.
Are your environments guaranteed to be deterministic and resistant to shortcuts?
No guarantee is made on this site. Those are properties an environment is specified against.
Whether a particular environment meets them is shown by its validation suite and calibration report, where an engagement includes them.
Agent evaluation and red teaming
What does agent evaluation involve?
Engineers assess coding and reasoning agents across correctness, architecture, security and failure modes, and experts judge model outputs against documented, calibrated rubrics.
Each failure comes with an expert diagnosis, not only a score.
Is Red Team the same as penetration testing?
No. Synthium Red Team is adversarial evaluation of model and agent behaviour: unsafe output, instruction override, data leakage and tool-use failures.
Conventional network, infrastructure and web-application penetration testing is not offered as a service on our Red Team page.
What does a red-team finding contain?
Each finding is severity-rated and ships with reproduction steps, so your team can verify it for itself.
Mitigation recommendations are part of the engagement. A retest after you make changes happens only where it is scoped.
Experts and quality
How are experts qualified?
Every contributor passes a project-relevant capability assessment, then completes project training and gold-standard calibration against your rubric before touching production data.
Screening is against the expertise a project actually requires, not a generic test.
What happens when expert reviewers disagree?
Disagreements are escalated to senior adjudication rather than averaged away or forced through a single pass.
Contributors are also scored continuously. Underperformers are recalibrated or removed, and consistent high performers are given more capacity.
What evidence of quality can you show?
We do not publish production quality numbers or customer names on this site. We work under NDA, and real scores belong to real engagements.
What we can share, in a pilot or under confidentiality, includes qualification and calibration pass rates, reviewer agreement, rework rates and turnaround times. This site shows the measurement system, not scores we have achieved.
Security, privacy and data handling
How is sensitive material handled?
Contributors get access only to the projects they work on. Client material is kept in separate, access-controlled project spaces, in private storage, and protected in transit over TLS.
For projects involving unreleased models or training data, a specific security plan is typically agreed up front and in writing before production work begins.
Can we work under an NDA?
Yes. Standard and custom NDAs are available on request, and contributors are bound by confidentiality terms before they see client material.
An NDA is a separate agreement for your engagement. It is not something these website terms provide.
Do you hold SOC 2, ISO 27001 or similar certifications?
This site does not claim any security certification.
If a certification or audit is a requirement of your engagement, we will scope it with you and tell you exactly where we stand.
What happens to our data when an engagement ends?
Retention periods and deletion schedules are defined contractually for each engagement. Client data is deleted on request or at the end of the engagement, per your instructions.
What personal information does this website collect?
Only what you enter into our application, enquiry and data-partnership forms, plus the technical details needed to run and protect the site. The Privacy Policy lists each form, the service providers involved, retention and how to make a request.
For enquiries, describe the work rather than sending the data.
Pilots and engagement
What is a pilot?
A pilot is a small, scoped piece of real work. We agree the task, the expert profile and the quality criteria, qualify and calibrate experts against your rubric, and produce a batch of output you can judge against those criteria.
No long-term commitment is required to start. Capacity is added only if the standard held on your work.
What do you need from us to scope work?
Your task, and what a good result looks like. Real examples help if you have them.
Later we need your rubric and guidelines, answers to edge-case questions as they come up, and your review of the first batch against the criteria agreed at scoping.
What happens after I send an enquiry?
We read it and usually reply within 1 business day with questions or a proposed pilot plan. A reply is not an offer or a contract.
Describe the work in words. Please do not paste samples, personal records, credentials or confidential files into the form.
How is work priced?
We do not publish prices, and this site does not state pricing. Commercial terms are agreed in writing for each engagement, after scoping.
Separate agreements, such as a statement of work or a data licence, cover them. The website terms do not.
How quickly can you start, and how much can you handle?
We do not publish timelines or capacity figures, because both depend on the expertise required and the calibration bar. They are agreed per project at scoping.
Throughput grows by adding qualified contributors, not by lowering the bar, and we scale only after the standard has held on your work.
Still deciding?
Send us the problem. We'll tell you how we'd approach it.
Or email us directly at hello@synthiumlabs.tech