#014 • 1 Apr 2025

Contents of this Newsletter

🧞‍♂️

Mastering QA Through Context

⏰ : 5 min

📖

Blog Highlights

⏰ : 10 min

💻

This Week in AI

⏰ : 3 min

🛠️

New Free Testing Tool

⏰ : 5 min

MAIN STORY

Would You Trust A QA Who Knows Nothing About Your App?

We've said it before, and we’ll say it again. Context is EVERYTHING for us. It’s not just a feature; it’s the backbone of how our AI understands your app. But what exactly is context? In simple terms, it's what gives meaning and direction. Without the right context, even the smartest AI can make mistakes.

A skilled QA engineer already knows how to write test cases, create scripts, and identify edge cases. But without deep knowledge of the app—its screens, flows, and interactions—they can’t accurately cover everything. The better their understanding of the application, the better their test coverage.

Large Language Models (LLMs) are trained on vast datasets, allowing them to recognize patterns and generate responses. But without specific context, their outputs remain generalized. Context is what transforms an AI from simply "generating a response" to understanding the task.

A lack of context can also lead to misinterpretations and biases. Early LLMs, for example, often defaulted to gender stereotypes—generating a male CEO and a female nurse. That’s because AI models inherit biases from the datasets they’re trained on. Without intervention, these biases shape their outputs. By injecting relevant context, we help override these biases, ensuring LLMs don’t just make predictions but actually understand the task at hand.

Context in Code Analysis

At Quash, our primary job is to collect and understand context. A good QA engineer doesn’t just run test cases—they learn the application inside out. They explore the app, navigate workflows, analyze past issues, and build a deep understanding of how everything fits together. This process takes time, often weeks or months.

Quash does the same, but instead of manually interacting with the app, our AI learns from the codebase, PRDs, existing workflows, analytics, crash reports, and figma files and integrated data sources. We push our systems to understand applications just as a seasoned QA would—by piecing together every available signal into a comprehensive picture.

1. Knowledge Graphs

We start by breaking down code into structured layers:

File Type – UI or logic files
Functionality – Login screens, registration flows, etc.
Description – A fine-grained breakdown of what each file does

This forms Context 1. We then feed this information to an LLM, instructing it to create entities and relationships which form Context 2, a knowledge graph. Next, we refine this graph to extract workflows, resulting in Context 3—our final, structured context. For the AI, this context is more than just data, it's a detailed understanding of the application. Covering screens and interactions and possible UI paths and workflows.

When code changes, this system helps us immediately identify affected workflows and generate test cases, making our testing process modular and adaptive.

2. Code Indexing

Indexing answers a simple but crucial question: Where is what? Think of how libraries organize books. Google’s early search engines worked similarly, structuring data to make it searchable. We apply the same principles to code analysis.

We embed files using an LLM, converting them into numerical representations stored in a vector database. From this, we create structured contexts:

Context 1 – Raw embeddings stored in a vector database
Context 2 – Navigational flow of the file in natural language
Context 3 – Screen activities in the file
Context 4 – Locator IDs for UI elements

With these contexts, we generate workflows and enriched workflows, which help us pinpoint affected areas when changes occur. This results in automatically written test cases and scripts that adapt to evolving codebases.

3. Reinforcement Learning for App Understanding

We're also implementing Reinforcement Learning (RL) to improve how our AI understands applications. RL enables AI systems to learn through trial and error, optimizing performance by maximizing rewards for correct actions. Some RL techniques we’re trying out include Monte Carlo methods and Q-learning. These approaches introduce adaptability, allowing our AI agents to refine their understanding dynamically, just as a human tester would.

4. Expanding Context Beyond Code

We don’t stop at just analyzing the code. Code is context, but so is everything else that influences how an application functions.

Analytics as context – Understanding user behavior and system interactions to detect anomalies.
Crash analytics (Crashlytics) as context – Identifying failure points and tracing them back to their root causes.
Documents and PRDs (Product Requirement Documents) as context – Mapping the intended workflows, feature expectations to better understand how the application is evolving and align with business goals through Notion and Confluence integrations.
API logs, version history, and real-time monitoring as context – Capturing every available data point to build a full understanding of the application
Figma files as context – Extracting design structure, UI and layouts intend to bridge the gap between what’s designed and what’s developed.

By aggregating all these contextual layers, Quash doesn't just test blindly. It learns, adapts, and tests intelligently, ensuring every insight contributes to better automation and more reliable software.

Why It Matters

A human QA engineer spends weeks understanding the app, workflows, and past issues. They build deep context over time, which allows them to make informed testing decisions.

Our LLM needs that same level of understanding. But instead of taking weeks, it gathers and processes this context in minutes. We don’t just run tests; we build an AI-driven understanding of your app.

Testing isn’t just about running checks, it’s about truly knowing your software.

And that’s what makes all the difference.

Check out our Recent Blogs

How to Building a Modern CI/CD Pipeline

A comprehensive guide on designing and implementing a modern...

Read article

QA vs QC: Key Differences Explained for Software Teams

QA and QC are often confused but serve distinct roles in the software...

Read article

news

This Week in AI

Qwen2.5-Omni: The New Multimodal Champion

Alibaba’s open-source Qwen2.5-Omni beats Gemini 1.5 Pro in multimodal reasoning, handling text, images, audio, and video in real time. Its Thinker-Talker design separates reasoning from speech, creating structured, natural responses. It even matches human-level speech quality.

Google’s Gemini 2.5 Pro: The Ultimate Thinker

DeepMind’s Gemini 2.5 Pro is in first place in reasoning, coding, and conversation, outpacing GPT-4.5 and Claude 3. With a 1M-token context (soon 2M!), it’s tackling long-form content, multimodal understanding, and agentic programming.

GPT-4o Image Generator: Internet’s New Obsession

The new GPT-4o image generator is so good it’s breaking the internet. Users are flooding social media with Studio Ghibli-inspired AI art, proving that AI-generated images can be both stunning and meme-worthy. At one point Sam Altman had to step in and tweet—

Sam Altman

@sama

it's super fun seeing people love images in chatgpt. but our GPUs are melting. we are going to temporarily introduce some rate limits while we work on making it more efficient. hopefully won't be long! chatgpt free tier will get 3 generations per day soon.

10:2 PM • Mar 27, 2025

3144

Retweets

69530

Likes

Read 2778 replies

Sam Altman

@sama

can yall please chill on generating images this is insane our team needs sleep

10:32 AM • Mar 30, 2025

6266

Retweets

125437

Likes

Read 6998 replies

Our team joined the fun, crafting whimsical Ghibli-style portraits that blend nostalgia with cutting-edge AI creativity. This trend has sparked discussions about creativity and copyright, highlighting the evolving landscape of AI in artistic expression. Would you like to see some of our Ghibli-inspired creations?

tools

Try Our PRD To Test Case Generator

We’ve been building something exciting—a new (and free!) tool to make test-driven development effortless. Our PRD to Test Case Generator, a key part of our larger New Feature Testing approach, helps teams generate test cases directly from PRDs and Figma designs before development even begins.

With this tool, you can:

Adopt Test Driven Development seamlessly by starting with test cases before development.
Reduce last-minute debugging and release with confidence.
Speed up collaboration between product, dev, and QA teams.

It’s live now, completely free to try, and we’d love for you to check it out!

Contents of this Newsletter

🧞‍♂️

Mastering QA Through Context

📖

Blog Highlights

💻

This Week in AI

🛠️

New Free Testing Tool

MAIN STORY

Would You Trust A QA Who Knows Nothing About Your App?

Context in Code Analysis

1. Knowledge Graphs

2. Code Indexing

3. Reinforcement Learning for App Understanding

4. Expanding Context Beyond Code

Why It Matters

How to Building a Modern CI/CD Pipeline

QA vs QC: Key Differences Explained for Software Teams

news

This Week in AI

​Qwen2.5-Omni: The New Multimodal Champion

​Google’s Gemini 2.5 Pro: The Ultimate Thinker

​GPT-4o Image Generator: Internet’s New Obsession

tools

Try Our PRD To Test Case Generator

Qwen2.5-Omni: The New Multimodal Champion

Google’s Gemini 2.5 Pro: The Ultimate Thinker

GPT-4o Image Generator: Internet’s New Obsession

Inside Our QA Revolution: Why Context Is King & A Free Tool Launch 🚀