Inspiration

Anyone who has worked in customer support knows the pain: the same 20 questions answered hundreds of times a week, response times stretching during peak hours, and customers left waiting when a single agent is handling five chats at once. We've lived this — both as support agents and as users stuck in a queue for a simple question.

When we saw how capable modern LLMs had become at grounded, conversational Q&A, we realized most support tickets don't actually need a human — they need context. That insight became Aiva: an AI customer support agent that resolves the repetitive 80% instantly and hands the critical 20% to humans, with full conversation context attached.

What it does

Aiva is a plug-and-play AI support agent:

  • Learns from your knowledge base — upload docs, FAQs, and past tickets; Aiva indexes them and answers with citations, not hallucinations
  • 24/7 instant responses — resolves common questions in seconds, in multiple languages
  • Smart human handoff — detects frustration, refund requests, and out-of-scope questions, then escalates to a human agent dashboard with the full conversation summary
  • Agent dashboard — teams review AI conversations, correct answers (which feed back into the knowledge base), and track resolution metrics

How we built it

We built Aiva with a spec-driven workflow in Kiro: before writing code, we generated structured requirements and a design document from a rough one-line idea, then let Kiro break it into implementable tasks. This kept the scope realistic for a hackathon and made every feature traceable to a requirement.

The stack: a Next.js + TypeScript frontend with a real-time chat widget, a Node.js backend, and Amazon Bedrock (Claude) as the LLM. Answers are grounded via RAG — documents are chunked, embedded, and stored in PostgreSQL with pgvector, with retrieval-augmented prompts enforcing citations. Live chat runs over WebSockets, and the escalation flow notifies human agents through the dashboard.

Challenges we ran into

  • Hallucination control: early prototypes confidently made up refund policies. We fixed this with strict grounding prompts, retrieval confidence thresholds, and an "I don't know — escalating to a human" fallback.
  • Knowing when to escalate: tuning the boundary between "AI should try" and "human needed now" took several iterations on sentiment and intent classification.
  • Latency: streaming responses and caching frequent retrievals brought perceived response time down to conversational speed.

What we learned

Working spec-first in Kiro changed how we build: writing requirements before code caught scope creep early and made iteration painless when we pivoted the handoff design mid-hackathon. Technically, we learned that production-grade AI support is less about prompt cleverness and more about retrieval quality, honest fallbacks, and a good human-in-the-loop experience.

What's next

Voice support, more channels (WhatsApp, Slack, email), automatic knowledge-base updates from corrected answers, and deeper analytics on what customers actually ask about.

Built With

Share this project:

Updates

Submission history