Skip to content
All work
Applied AI2024

AI Chat Application

A transformer-based assistant that turns a support backlog into a real-time conversation, with inference fast enough to feel like a person is typing.

Role
Product & Engineering Lead
Stack
PythonAI/MLNLP

Problem

Support conversations were queued, not answered. Every question — including the eighty percent that were variations on the same dozen intents — waited behind a human. The cost curve rose with volume while the experience got worse, and no amount of hiring changed the shape of that curve.

Solution

We built a transformer-based conversational layer that resolves the repetitive majority on its own and hands the genuinely novel cases to a human with full context attached. The design decision that mattered was scope: rather than chasing a system that answers everything, we defined a confident subset, measured it honestly, and made escalation a first-class path instead of a failure state. Real-time NLP inference kept latency inside the window where a conversation still feels like a conversation.

Outcome

Response time collapsed from queue-length to sub-second for the covered intents. Human effort shifted from triage to the exceptions that actually needed judgment, and the volume-to-headcount relationship stopped being linear.

Metrics

Bracketed figures are placeholders pending final reporting.

reduction in response time
[XX]%reduction in response time
of conversations resolved without escalation
[XX]%of conversations resolved without escalation
increase in support throughput
[XX]%increase in support throughput

Approach

The decisions that shaped it.

  1. Framed the intent taxonomy from real transcripts before writing any model code.
  2. Set a confidence threshold that favored graceful escalation over confident errors.
  3. Optimized the inference path for perceived latency, not just raw throughput.
  4. Instrumented every escalation as a labeled training signal for the next iteration.