(See also: Implementation, Our Approach)
ZTS provides advisory and education services for organizations evaluating, deploying, or governing LLM systems.
This work is informed by direct involvement in AI safety research and technical education, including participation in the AI Mental Health Collective and The Human Line Project.
Consulting
We provide consulting for organizations evaluating LLM technology, planning implementations, or developing their approach to AI adoption.
Technology Evaluation: Assessment of commercial APIs, open-source models, and deployment approaches for specific use cases. We help determine what’s actually feasible, what claims don’t hold up under scrutiny, and where AI creates value versus where it introduces unnecessary complexity.
Architecture Review: Analysis of proposed or existing LLM implementations, covering reliability, security, cost structure, and maintainability. Independent review before committing resources.
Readiness Evaluation: Assessment of organizational preparedness for AI adoption, covering technical infrastructure, staff capabilities, governance requirements, and cultural factors.
Implementation Planning: Guidance on sequencing, resource requirements, and risk management for AI initiatives. We help define requirements, identify pitfalls, and establish realistic timelines and expectations.
AI Safety & Governance
ZTS approaches AI deployment with attention to safety, ethics, and organizational control. Organizations need policies that define acceptable use, oversight mechanisms that provide meaningful control, and realistic assessment of what can go wrong.
We work with organizations to develop governance frameworks covering how AI systems should be used, what technical and procedural controls should be in place, and how to maintain human oversight that isn’t just performative. This includes developing usage policies, conducting risk assessments, and providing guidance on implementing guardrails and content classification systems.
Work focuses on helping organizations evaluate risks, establish clear policies for AI use, and implement systems aligned with organizational values and requirements. We help ensure systems remain transparent and controllable.
Our approach to safety comes from studying how models actually behave in practice, not from compliance checklists. We focus on controls that address real failure modes rather than hypothetical concerns.
Focus areas:
- AI usage policy development tailored to organizational context
- Risk assessment covering model behavior, data handling, and operational failures
- Guardrail implementation including classifier-based content filtering
- Documentation for compliance and internal accountability
- Incident response planning for AI system failures
Model Evaluation & Red Teaming
Before deploying an LLM system, particularly in customer-facing or high-stakes contexts, organizations benefit from understanding how the model actually behaves under varied conditions. Vendor documentation and benchmarks rarely capture the failure modes that matter in production.
ZTS provides adversarial testing and evaluation designed to surface problems before deployment. This includes structured red teaming sessions that probe for jailbreaks and policy violations, robustness assessments that test behavior at edge cases, and comparative evaluations that cut through marketing claims.
This work builds on experience analyzing models from multiple providers, with attention to the gap between stated capabilities and operational reality. We’ve published analysis on bias, censorship, and safety issues in commercial models, and that critical perspective informs our evaluation work.
Evaluation services:
- Adversarial prompting to test jailbreak resistance and policy enforcement
- Bias and fairness assessment across demographic categories and use cases
- Safety evaluation for harmful content generation and misuse potential
- Guardrail and filter effectiveness testing under adversarial conditions
- Comparative evaluation to support model selection decisions
Technical Training & Workshops
Most organizations adopting LLM technology lack the internal expertise to evaluate what these systems actually do, where they fail, and how to use them effectively. Training resources available online often reinforce misconceptions, repeat marketing language, or skip the fundamentals that matter for practical work.
ZTS provides technical training focused on practical implementation. Training emphasizes production considerations, covering prompt engineering, application architecture, cost management, and building internal capabilities. We cover how these systems actually work at an architectural level, what they can and cannot do, and how to interact with them effectively. Training avoids the anthropomorphization and hype that clouds most AI education.
Content is adjusted based on audience. Training for developers focuses on implementation details and failure modes. Training for executives focuses on capabilities, limitations, and organizational implications. Mixed audiences get a practical foundation that supports cross-functional conversations.
Training topics:
- LLM architecture: how transformers work, what training and fine-tuning do, inference mechanics
- Prompt engineering: practical techniques, common failures, and why “prompt hacking” articles are mostly wrong
- Application architecture and cost management for LLM systems
- Output evaluation: verification strategies, hallucination detection, building appropriate trust
- AI literacy for leadership: realistic capability assessment, risk understanding, decision frameworks
- Responsible deployment: safety considerations, user communication, maintaining human oversight
Technical Content Development
Organizations deploying AI systems often need internal documentation, training materials, and policy documents that address their specific context. Generic resources don’t account for particular tools, workflows, or organizational requirements.
We develop custom technical content including internal usage guides, training materials, governance documentation, and explanatory resources. This work applies experience translating complex technical concepts into clear guidance that people actually use.
Documentation is written for the intended audience, whether that’s developers integrating an API, support staff using an AI-assisted tool, or executives evaluating strategic decisions. We focus on actionable guidance rather than comprehensive reference material.
Content development:
- Internal AI usage guides covering approved tools, best practices, and limitations
- Training materials including slide decks, exercises, and reference documents
- Policy and governance documentation suitable for internal use and external review
- Technical explainers that give non-technical stakeholders accurate understanding
- Scenario-based playbooks covering common situations and edge cases