hello.harshweb@gmail.com
Back home

Aria

An AI support agent that answers most tickets before a human ever sees them.

Client

Aria Systems

Services

AI Integration,
Full-Stack Development

Type

AI Product

Aria — main screen
Aria — main screen
Aria — main screen

Introduction

Aria's support team was drowning. The same fifteen questions made up most of the queue, and first-response time was measured in days, not minutes.

I built an AI agent that answers from their own documentation, escalates when it is unsure, and shows its work in an admin panel the team actually controls.

Goals

  • Cut first-response time to under a minute
  • Keep every answer grounded in real documentation
  • Escalate to a human gracefully, with context
  • Give the team visibility into accuracy and gaps
  • Ship inside their existing helpdesk workflow

Target audience

  • Support agents handling 200+ tickets a day
  • Customers expecting an instant answer
  • The ops lead responsible for tone and accuracy

Approach

Retrieval first, model second. The agent searches their docs, cites the source, and refuses to answer when confidence is low — which is exactly what makes it trustworthy in production.

An eval harness runs a set of real historical questions on every prompt change, so improvements are measured instead of guessed.

The interface streams answers token by token so the product feels fast even when the model is thinking.

Design

A chat surface that inherits the host site's typography

Confidence indicator and source chips under every answer

Admin dashboard: transcripts, gaps, escalation rules

Deliverables

  • Chat widget and embeddable API
  • Retrieval pipeline over their documentation
  • Admin dashboard and analytics
  • Evaluation harness and prompt versioning
  • Deployment, monitoring and handover docs

Stack

Next.js, OpenAI, Vector search, Node.js, Redis, Vercel

Next project Cadence — startup website