DrizzVision bets on visual AI to test mobile apps
The Indian startup DrizzVision aims to make mobile testing less fragile with Fathom, an agent that reads the screen like a human. Its own measurements show very variable performance depending on the environment.
In Bengaluru, DrizzVision aims to simplify a profession where the slightest interface tweak can derail a testing campaign. Its agent Fathom uses visual models to understand the screen of a mobile application and execute instructions in natural language, without relying solely on fixed technical identifiers. The idea matters for product and QA teams because it targets tests that break at the first change of a button, page, or pop-up. One stubborn point remains: reliability drops as soon as the environment changes.
When the screen changes, the test breaks less quickly
The project was born from a common frustration in mobile development: moving fast then spending two or three days on end-to-end testing, according to Yash, one of the co-founders. With Asad, he began working on this idea in 2024, when vision models became capable of interpreting computer screens. DrizzVision then moved away from traditional automation tools, like Appium, which rely on the internal structure of an application to click on specific elements.
This method works as long as the interface remains stable. But a renamed button, a redesigned page, or an unexpected window is enough to break the script, even if the business function hasn’t changed. DrizzVision says it wants to “just understand the screen and execute,” in other words recognize what is visible on the screen and act accordingly. A simple promise on paper. Much less so on a real phone.
Pop-ups, icons and “self heal”: the real playground
The startup takes the example of a food delivery app. During an order, a user may encounter a promo offer, a permission request, or another pop-up window. The agent is supposed to ignore these interruptions and continue the planned flow. “A pop-up ideally should not break your test case,” said Yash. “It should say that your checkout functionality is working fine. Although there were three pop-ups that we saw on our way.”
The same reasoning applies to interface changes. If an “Edit” button becomes a pencil icon, the AI should still understand the intended function and continue. Yash talks about a system that “self heals,” meaning it adjusts its behavior when the interface moves. DrizzVision claims to have reduced “test flakiness,” that is the rate of unstable or intermittent tests, from 15% to 50% down to about 3% on its internal benchmark, with results varying depending on the client and use case.
Very uneven results depending on models and screens
The figures provided by DrizzVision invite caution, however. The company tested Fathom on Android with six frontier model configurations and 20 tasks, in two distinct environments, repeating each combination five times. In the simplest suite, on a standard Android image, the models succeeded in 217 scored runs out of 300, or 72.3%. In the more difficult suite, which used a rooted Android image and more complex tasks, they succeeded only 86 times out of 300, or 28.7%.
The differences between models are also marked. GPT 5.6 Luna scored 64% on the simplest suite but 4% on the hardest. Gemini 3.5 Flash performed worse on the first, with 54%, then better on the second, with 34%, ahead of all GPT configurations tested on this aspect. DrizzVision specifies that this benchmark measures the entire system — model, Fathom agent, and Android image used — not the model alone. In other words, comparing these scores as if comparing 100-meter sprint times would be a bit too comfortable.
Data security and the emerging market
The question is not only about performance but also about data. “The screenshot should not leave the device,” explained Yash, while the tested applications may contain personal or financial information. For its client deployments, DrizzVision says it uses its own on-premise models rather than relying entirely on large cloud models. The co-founder also mentions as a near-term goal a model with fewer than one billion parameters capable of handling half or more of the workload.
Commercially, the startup began with a command-line tool, piloted with Walmart, where users wrote tests in text files to run them on local emulators. The product is now accessible via desktop, web, CI/CD pipelines, and an MCP interface, this protocol allowing code assistants to control external tools. DrizzVision claims more than 10 enterprise clients and over 20 SMEs in India, as well as permission to name Tata 1mg among them. It raised $2.7 million in seed funding and employs about 15 people.
The company also faces tools like Maestro, widely used in the United States. Yash expects DrizzVision to complement Maestro more often than replace it, as the two products do not target exactly the same problem: Maestro relies on YAML flows, automatic waits, and the accessibility tree, while DrizzVision targets tests that break when the interface changes. The startup is already looking further ahead, towards verifying AI agents capable of acting on behalf of users, a market it estimates potentially seven to ten times larger than its current contracts. It is not yet revenue, but the opportunity is on the table.
Key points
- DrizzVision was founded in Bengaluru in 2024.
- Fathom reads the screen to execute mobile tests.
- The internal benchmark dropped from about 15-50% to 3%.
- 72.3% success rate on the simplest Android suite.
- 28.7% success rate on the most difficult Android suite.
- $2.7 million raised in seed funding.
- Tata 1mg is cited as an authorized client.
By the numbers
- 20 tasks — tested on Android by DrizzVision.
- 300 scored runs — per benchmark suite, across six models.
- More than 10 enterprise clients — in India, according to DrizzVision.
- More than 20 SME clients — in India, according to DrizzVision.
- Less than 1 billion parameters — target mentioned for an internal model.