Liina Aava

Designed and operated a production RAG platform using dense retrieval, BM25, metadata filtering, and reciprocal rank fusion to improve answer relevance and reduce hallucinatory outputs in enterprise workflows. Built an agentic AI system with structured tool orchestration (SQL, policy lookup, web search), implementing validation, retrievers, circuit breakers, and loop-prevention safeguards. Implemented scalable ingestion and indexing pipelines covering document normalization, chunking, deduplication, drift detection, and re-indexing across multiple data sources. Introduced evaluation and monitoring frameworks with grounding, latency, and cost-SLIs, integrating automated checks into CI/CD to prevent regressions. Established prompt lifecycle management with versioning, tenant-specific templates, and adversarial testing against prompt injection and jail break attempts. Optimized latency and infrastructure costs through semantic caching, batching, and dynamic model routing with auto-scaling

Built a compliant Q&A assistant using LLMs and retrieval pipelines to surface contractual clauses and regulatory references for audit workflows. Implemented document ingestion and retriever systems with change detection, hashing, FAISS re-indexing, and Airflow/backfills for large contractor repositories. Designed LLM safety guardrails including input sanitization, allow/deny lists, sandboxed tool execution, and strict timeout controls. Established telemetries and evaluation frameworks capturing latency, token usage, model versions, and rubric-based quality checks to prevent unsafe production changes. Supported regulatory audits with reproducible inference logs, signed outputs, and versioned artifacts enabling traceable decision re-execution.