All work
AI / LLMApr 2024 – Sep 2025

GPT-powered self-healing for automation

When selectors fail and alternate IDs don't match, LLM-powered label suggestions find the right element — demonstrated live on a website from 2013.

Featured on Microsoft Developer YouTube · Bridges full UI redesigns


The problem

ACCELQ's alternate-ID self-healing algorithm covers most UI changes automatically. But when the change is large enough — a completely redesigned layout, renamed elements, restructured components — even alternate IDs fail. A human had to investigate and manually fix the broken step. For enterprises running hundreds of tests against fast-moving products, this was a significant maintenance burden.

What I built

Integrated GPT into locator-free troubleshooting as a last-resort layer. Rather than sending the full page DOM to the LLM (expensive and noisy), the system pre-calculates possible element matches by type and sends only the relevant labels to GPT. GPT then suggests the most likely correct element.

Also implemented locator-free element type conversion — when an element changes type (e.g. a button becomes a link) but has the same or matching label, it converts automatically without any GPT call.

Created a real-world demo on the Southwest Airlines website from 2013 (via Wayback Machine) vs the current version — demonstrating decade-long UI drift being healed automatically in real time. This demo was featured on the Microsoft Developer YouTube channel.

Approaches considered

Send full page DOM to GPT — rejected: expensive per token, too much noise, GPT hallucinates on large unstructured HTML input
Computer vision / screenshot matching — considered: slower, requires significant additional infrastructure
Vector embeddings for element matching — considered: good for static sites, doesn't handle element type changes well
Pre-filtered label set by element type + GPT — chosen: reduces token cost, focuses GPT on the right candidates, dramatically improves accuracy

Architecture

Primary selector fails → alternate-ID healing attempted → if still failing → element type detection → label pre-computation for that element type only → minimal-context GPT request → suggestion applied → test continues. GPT is a last resort, not a first resort; the fast algorithmic layers handle the majority of cases with zero LLM cost.

Result

Self-healing now covers the full spectrum: minor UI changes (alternate-ID), moderate changes (locator-free matching), and major redesigns (GPT suggestion). The Southwest Airlines 2013 demo — showing a working test against a 10-year-old page layout healing itself to the current site — was featured on the Microsoft Developer YouTube channel.

GPT / LLMJavaScriptSelf-HealingLocator-Free
Back to all work