A QA engineer at a brokerage explains how they integrated large models into a shift-left process to speed requirements review, expand coverage, and keep human judgement where it matters.
Use an LLM as a partner during requirements analysis to surface gaps, ambiguous acceptance criteria, and unhandled edge cases before development starts.
Combine a standard test stack (Python/pytest, Playwright, OpenCV, EasyOCR, PyAutoGUI) with a context layer (MCP) that supplies internal rules and docs to the model rather than fine-tuning it.
Keep human control gates: AI automates routine drafting and checks, but testers retain final decisions, exploratory testing, and acceptance calls.
# Background Dzianis Talstsiuk, a QA engineer at EXANTE, describes how he built AI into testing for desktop, web and mobile trading terminals. The aim: speed up routine work, broaden test coverage, and free testers to focus on analysis and higher-value tasks. The domain is financial trading, where rendering or state errors can cause client losses and regulatory risk, so requirements and acceptance criteria must be precise.
# Tools and technical stack The project keeps a conventional automation stack and layers generative models on top:
Rather than fine-tuning, the team provides a context layer (called a Model Context Protocol, MCP) that connects the model to internal documentation, testing guides, and prior feedback. This gives the model project-specific context when reviewing requirements and generating test artefacts.
# How the workflow changed Traditional flow: QA often joins late and reacts to implementation gaps found after code is written. That creates rework and slower delivery.
New flow: QA joins as soon as requirements are fixed by Product. The tester runs the requirements through the LLM to check for completeness, ambiguity, logical gaps, and missing edge cases. The model returns structured questions and drafts that the tester reviews and refines.
Concrete example: a feature to pin a focused instrument at the top of a list appeared complete on first read, but the model flagged a missing case for the search screen. That gap, if found after implementation, would require rework. The model helped catch it earlier.
# What AI does well in this setup
# Where AI falls short
# Practical implementation notes
# Outcome and recommended approach Embed an assistant at the QA entry point to shift testing left in a way that preserves human control. Use models to reduce routine drafting and to act as a fresh pair of eyes during requirements review. Keep explicit control points where testers make the final calls about acceptance and release readiness. Where domain risk is high, treat model output as a structured aid rather than an automated decision.
AI-native test planning uses AI to draft cases from requirements and rank coverage by risk. Learn the workflow, the review gate, and what AI cannot decide.
Traditional test automation frameworks often carry heavy maintenance costs, slow release cycles, and high knowledge dependency. By transitioning from standard script creation to a governed AI Test Automation Factory , engineering teams can shift their focus from writing boilerplate code to high-value validation and arc
Just a few months back, I observed a test suite with a self-healing feature “fixing” a failed selector three different times during the same sprint cycle. In each instance, the fix performed its function perfectly well; however, it didn’t address the real problem of a shipped UI regression, as its sole focus was on kee
Originally appeared on Evil Martians . Authors: Yuri Mikhin, Frontend Engineer, and Travis Turner, Tech Editor Topics: DX, AI, Agent Experience, TypeScript, React, Agentic development, LLMs, Continuous Integration Ten checks that catch what AI-written frontend code hides: contract codegen, boundary linting, mutation te
Learn how Spec-Driven Development uses specifications, technical planning, and AI-assisted development to reduce ambiguity, rework, and implementation risk.
Loading more related stories...
Open the app view to save this story, compare related coverage, and continue from the same source.