Course Objective & Scope
Having individual application teams write their own PII scrubbing logic leads to inconsistencies and leaked personal data. This course focuses on engineering a centralized PII gateway in front of LLM APIs that accurately detects and masks personal identifiers across Turkish and English prompts.
In hands-on labs participants build hybrid recognisers combining Microsoft Presidio, GLiNER and national ID/IBAN checksum validators, and deploy a streaming-compatible gateway service backed by Redis/Vault session mapping that restores typed placeholders ([PERSON_1]) in model responses.
Who Should Attend?
- Backend and AI engineers building AI gateways and platform services
- Data scientists working with NLP and Named Entity Recognition (NER) models
- Engineers responsible for data security and DLP integrations
- Prerequisite: experience with Python (FastAPI/async) and REST API development
Key Takeaways
- Write custom Presidio recognisers for Turkish national/tax IDs, IBANs, phone numbers and internal customer IDs
- Combine GLiNER and Turkish transformer NER models with deterministic rules to maximise recall
- Implement reversible masking with typed placeholders that preserve semantic coherence for the LLM
- Build a streaming-aware de-masking layer on top of LiteLLM or a custom FastAPI proxy
- Benchmark gateway performance using precision, recall, F2 score and p95 latency
Curriculum
Designing a Hybrid PII Detection Engine
- Rule-based detection with regex, checksum validation (TCKN/Luhn/IBAN) and context words
- Integrating Microsoft Presidio, spaCy/Hugging Face Turkish NER and zero-shot GLiNER
Masking Strategies and Semantic Coherence
- Comparing redaction, hashing, format-preserving encryption (FPE) and typed placeholders (
[CUSTOMER_1]) - Consistent entity-to-placeholder mapping across multi-turn conversations
Reversible Flows in a Proxy Architecture
- Managing ephemeral mapping tables with TTLs and encrypted Redis/Vault storage
- Buffering and restoring split placeholder tokens across SSE streaming responses
PII Scrubbing in Multimodal Inputs (PDFs, Images, Transcripts)
- Post-OCR PII redaction in RAG ingestion and file upload workflows
- Bidirectional PII inspection on tool calling arguments and error payloads
Quality Benchmarking and Latency Optimisation
- Evaluating precision, recall and F2 on a synthetic Turkish PII benchmark
- Reducing gateway latency to milliseconds with ONNX Runtime quantization and caching
Other Courses in This Category
-
Teaches how to replace production customer data in testing, analytics and LLM evaluation environments using k-anonymity, differential privacy and SDV/LLM-based synthetic data generation.
- Live Course
- 5 modules
-
Teaches how to scrub personal data from LLM and agent telemetry in OpenTelemetry, Langfuse and SIEM logs, and engineer retention policies, access controls and KVKK compliance into observability pipelines.
- Live Course
- 5 modules
-
Covers Agent Goal Hijack—the top OWASP Agentic AI risk—showing how attackers divert autonomous planning loops, persistent memory and multi-step execution, and how to defend against it with intent verification.
- Live Course
- 5 modules
-
Covers tool poisoning, rug pulls, tool shadowing and supply chain risks in Model Context Protocol (MCP) servers, third-party agent tools and A2A integrations, and how to build a secure MCP architecture.
- Live Course
- 5 modules
-
Teaches how to prevent remote code execution (RCE) and injection vulnerabilities caused by executing LLM- and agent-generated Python, SQL, shell or template outputs using AST validation and safe output handling.
- Live Course
- 5 modules
-
Teaches how to isolate code-executing, file-processing and browser-using AI agents inside secure sandboxes using Firecracker microVMs, gVisor, Kata Containers, WebAssembly and strict network egress controls.
- Live Course
- 5 modules
Get in touch
