// Channels
Each desk routes to a named engineer with a published SLA. We do not staff a generic sales inbox; the wrong queue gets forwarded the same hour, not absorbed.
For Fortune 500 and scaling enterprises moving an AI agent into a business-critical workflow. We scope evaluation suites, red-team coverage, and drift gates around your stack.
contact@testml.orgTargeted prompt-injection, exfiltration, and role-collapse engagements. Bring a model name, a guardrail spec, and a threat surface. We scope a corpus and report against it.
redteam@testml.orgAuditor briefings, evidence packages, and BAA paperwork for regulated workflows. We map evaluator outputs to your control framework so review cycles stop blocking go-live.
compliance@testml.orgReporters, industry analysts, and conference programmers. We ship technical primers and methodology notes; we do not field hype-cycle commentary or generic AI predictions.
press@testml.org// What happens next
We work production-first. The pilot output is a working evaluator, not a slide deck — and the engagement ends when your engineers are running it without us.
A short intake describing the model, the workflow, the regulated surface, and what go-live decision the evaluation has to support. Three paragraphs is enough.
Same-day acknowledgement
An engineer drafts a domain-specific evaluation plan: accuracy, latency, jailbreak resistance, hallucination rate, and the regression gates we will run on every deploy.
Within 5 business days
Replayable runs against a versioned prompt-and-policy snapshot. You see the dashboards we use internally — not a marketing PDF — with raw traces attached to every claim.
Pilot window: 2–3 weeks
Methodology, evaluators, and drift monitors transfer to your team. We co-own the first regression cycle, then step back. No staffing dependency, no vendor lock-in.
Methodology transferred
// Send-us-this
Free-form prose is fine. We just need these signals so the on-call engineer can scope a corpus before the first call.
Vendor and version, or open-weights checkpoint hash.
Claude Sonnet 4.5, GPT-4o-2024-08-06, Llama 3.1-70B-Instruct, etc.
What the agent actually decides or produces.
e.g. claims triage, contract clause extraction, clinician-facing summary.
Regulatory surface and data sensitivity class.
Legal, medical, financial, defense, EU consumer, etc.
What a wrong answer costs in production.
Reversibility, blast radius, compliance penalty, auditor scrutiny.
Decision date the evaluation has to support.
Pilot, phased rollout, board-level go-no-go review.
// Before you write
No. Customer artifacts live in tenant-isolated storage with zero data retention by default. We do not train, fine-tune, or share corpora across engagements. Retention windows and BAA paperwork are negotiated per workflow.
Yes — production-first is the point. We instrument drift monitors and regression gates against the workflow you already have running, then back-fill an evaluation suite around the failure modes the telemetry surfaces.
No. Our methodology was built around multi-agent orchestration: routers, tool-callers, retrieval layers, and policy fences. We evaluate the system, not just the underlying model card.
We are not a staffing agency. Engagements transfer methodology, evaluators, and tooling to your engineers. The success criteria is your team running the regression cycle without us by the second quarter.
Download the standalone template from the Research index and run it yourself. If the output surfaces gaps you want a second opinion on, send us the filled-in brief and we will scope from there.