Course Objective & Scope
Writing 'follow the rules and never mention competitors' inside a system prompt is not a security control; a determined user or an indirect prompt injection hidden in a retrieved document can bypass it easily. This course teaches how to place deterministic and model-based guardrails in front of the model (input rails), after the model (output rails) and around tool calls (execution rails).
Participants work with Microsoft Presidio, Llama Guard, NeMo Guardrails and Guardrails AI to implement Turkish and English PII redaction, jailbreak detection, topic containment and schema/grounding checks within a tight latency budget.
Who Should Attend?
- AI and software engineers building customer-facing or internal LLM applications
- AppSec and DevSecOps teams responsible for AI safety, KVKK/GDPR compliance and risk controls
- Architects managing enterprise AI gateways and platform layers
- Prerequisite: Python and foundational LLM application architecture knowledge
Key Takeaways
- Filter direct and indirect (RAG-borne) prompt injection attempts at the input layer
- Redact personal data (PII) before prompts reach the model and handle re-hydration safely
- Constrain assistants to approved business topics and brand guidelines
- Catch ungrounded claims, competitor mentions and sensitive data leaks in model outputs
- Design parallel guardrail execution that preserves low latency and streaming UX
Curriculum
Defense-in-Depth Guardrail Architecture
- Checkpoints across input, retrieval, tool-execution and output layers
- Combining deterministic rules, fast BERT/SLM classifiers and LLM-based guards
Input Rails: PII Redaction and Prompt Injection Shields
- Turkish and English PII anonymisation with Presidio and custom regex/NER rules
- Detecting jailbreaks and indirect prompt injections (Llama Prompt Guard, DeBERTa classifiers)
Topic Containment and Dialogue Rails
- Blocking off-topic queries using NeMo Guardrails and embedding-based topic routers
- Enforcing legal/medical/financial advice boundaries and competitor filters
Output and Tool-Execution Validation
- Schema validation, context grounding checks and hallucination prevention
- Parameter bounds, permission checks and human-in-the-loop approval for high-risk agent tools
Latency Optimization, Streaming and False-Positive Tuning
- Running input guardrails in parallel with speculative LLM generation to protect TTFT
- Chunk-by-chunk validation on streaming responses and monitoring guardrail block rates
Other Courses in This Category
-
Cutting cost and latency in production LLM applications with multi-layer caching: provider and vLLM prompt caching, Redis exact caching, vector-based semantic caching, tenant isolation and cache invalidation.
- Live Course
- 5 modules
-
End-to-end latency management in multi-step LLM, RAG and agent applications: SSE streaming UX, parallel retrieval/guardrail/tool execution, output token reduction, P95/P99 tail-latency control and hedged request patterns.
- Live Course
- 5 modules
-
Incident management tailored to production LLM and agent systems: response runbooks for hallucination spikes, prompt injection/data leaks, runaway agent loops and provider outages, kill-switch controls and AI post-mortems.
- Live Course
- 5 modules
-
Combining managed cloud AI platforms (Azure AI Foundry, Amazon Bedrock, Google Vertex AI) with on-premises GPU clusters in a hybrid architecture: sensitivity-based routing, portable contracts, private networking and disaster recovery.
- Live Course
- 5 modules
-
Moving from ad-hoc 'vibe checks' to measurable quality in LLM applications: building a first eval dataset, rule-based assertions, basic LLM-as-a-judge grading and running a first automated evaluation report with Promptfoo.
- Live Course
- 5 modules
-
Designing high-quality golden datasets that anchor LLM evaluations: scenario matrices, synthetic test generation and filtering from documents, adversarial/edge-case design, expert annotation protocols and dataset maintenance.
- Live Course
- 5 modules
Get in touch
