The Ekho Navigator team developed an AI-powered student guidance assistant for CSU Channel Islands that routes students, families, and prospective students to campus answers about registration, advising, financial aid, tutoring, and locations through structured decision-making where the model classifies needs into one of four actions rather than generating free-form text. The solution is built using a React and Vite frontend with streaming markdown replies, backed by AWS Lambda (serverless compute) orchestrator functions behind Amazon API Gateway that implement a safety-first architecture with deterministic crisis detection before any AI processing. Campus resource information scraped from CSUCI websites is stored in Amazon S3 (cloud storage) as Markdown documents with metadata sidecars, feeding into Amazon Bedrock Knowledge Bases (managed RAG) using Amazon Titan Text Embeddings to create vector representations stored in Amazon OpenSearch Serverless (search and vector index). When a student submits a question, the Lambda orchestrator first runs a deterministic safety filter that matches crisis keywords and immediately returns a canned response with 988 Lifeline and campus counseling contacts, bypassing all AI processing. For Spanish mode queries, Amazon Bedrock (managed generative AI) using Claude Haiku models translates the question to English before querying the Knowledge Base to maximize match quality, then forces the final reply to be fully Spanish. The orchestrator retrieves the top five passages from the Knowledge Base and sends them to Amazon Bedrock Converse API using Claude Sonnet 5 models with a tool configuration that forces the model to call exactly one of four tools—answer with numbered passage citations, ask a clarifying question, escalate to a specific campus office, or decline as out of scope. The model cites passages by number only, never emitting URLs, and code resolves the numbers to trusted source URLs from retrieved chunks, with invalid citations causing fail-safe escalation to a human. When the model escalates, code assembles an escalation card with verified contact details from an office directory covering financial aid, advising, registrar, tutoring, and general support, along with buttons to generate a support ticket with a staff-facing summary. The Lambda orchestrator validates every tool decision for structural correctness, with any unusable model output triggering fail-safe escalation and infrastructure errors returning retryable responses, while Amazon Bedrock Guardrails applies content filtering, PII masking, and grounding checks, and Amazon SNS (notifications and SMS) handles escalation notifications, providing CSU Channel Islands with a student support channel that routes directly to the correct service with grounded citations and deterministic safety handling for crisis situations.