AI Customer Support & Agent Automation at Cresta

Overview

A client engagement applying generative AI to high-volume customer service. The system has two sides. The AI handles repetitive interactions on its own, and human agents get live suggestions, the right context, and a clean way to take over when a conversation really needs a person.

What I built

• AI support workflows that interpret customer intent, generate contextual responses, and handle common support scenarios end to end.
• Real-time Agent Assist surfacing suggested responses and relevant context to human agents while they're mid-conversation with a customer.
• Knowledge retrieval and RAG workflows that search approved company knowledge sources and ground the model's response in what it finds, rather than in what it remembers.
• Conversation summarization and structured extraction, turning long interactions into concise summaries with detected intents, actions, and follow-up items.
• Support analytics and escalation workflows tracking conversations, AI-resolved versus human-handled cases, recurring issues, and the situations that need to reach a person.

Technologies

• Frontend: React, Next.js, TypeScript, real-time conversation interfaces, support dashboards, agent-assistance UI
• Backend: Python, REST APIs, PostgreSQL, Redis, async and background processing, CRM and contact-center integrations, Docker, cloud infrastructure, logging and observability
• AI: LLMs, retrieval-augmented generation, vector search and embeddings, intent classification, conversation summarization, structured outputs, prompt and workflow orchestration, context management, confidence-based escalation

Impact

Automating the repetitive end of the support queue meant agents spent less time on interactions that never needed judgment, and Agent Assist cut down the manual searching between a customer's question and a good answer. Because every AI response was grounded in approved knowledge sources, answers were more consistent than when each agent phrased things their own way. Automatic summaries took over a good share of the after-conversation paperwork. Confidence-based escalation sent complex or uncertain cases to a person instead of letting the model guess, and the analytics made recurring issues and conversation trends easy to spot.

I built this with product managers, designers, AI/ML engineers, frontend and backend engineers, data engineers, QA, platform/infrastructure engineers, and customer-support domain stakeholders.

This work was completed as part of a broader product and engineering effort. The case study focuses on the customer-support, AI, and full-stack systems I contributed to rather than claiming sole authorship of the complete Cresta platform.