Polylingual AI Support Agent

Polylingual replaces ungrounded chatbot responses with an auditable decision pipeline. Every ticket incoming via Slack, Zendesk, or Email undergoes semantic retrieval, confidence scoring, and explicit human-in-the-loop fallback.

• Automatic language detection preserving customer language (French/English).
• pgvector similarity search combined with BM25 lexical ranker.
• NVIDIA NIM GLM 5.2 schema-validated model round trips.
• Real-time execution cost tracking and latency telemetry.

Traditional LLM support bots frequently make up policy details, quote wrong return windows, or provide contradictory advice. When customer service bots guess, support costs increase due to secondary escalations.

Confidence is not an answer. By treating low confidence as a first-class routing event rather than an edge case, we can achieve high automation rates without sacrificing accuracy.

Iterated through vector-only search, graph RAG, and hybrid rankers before settling on a lightweight dual-rank pipeline that delivers sub-second retrieval accuracy.

Artifact 01: Operations Overview & Metrics

Artifact 02: Case Intake Terminal

The CloudFront CDN routes API requests to ECS Fargate backend. Vector embeddings are generated via sentence-transformers and queried against pgvector. GLM 5.2 validates responses against exact source documents.

The core Fast-API resolution engine enforces atomic execution bounds, logging trace IDs and token metrics to Prometheus.

def evaluate_ticket_confidence(retrieved_docs, score_threshold=0.78):
    top_score = max([doc.score for doc in retrieved_docs]) if retrieved_docs else 0.0
    if top_score < score_threshold:
        return Decision(action="ESCALATE_TO_HUMAN", confidence=top_score)
    return Decision(action="AUTO_RESOLVE", confidence=top_score)

If the NVIDIA NIM endpoint experiences latency exceeding 35 seconds, the engine gracefully defaults to a grounded human-handoff payload with pre-populated context notes.

Achieved 88% auto-resolution rate on routine tickets with 0% unsourced claims in production evaluations.

Building for production RAG is 20% prompt design and 80% data hygiene, strict schema validation, and operational fallback engineering.