July 2021 to present
AI / Machine Learning Engineer
TestGrid Labs Inc.
CoTester, the flagship agent
- Core engineer on CoTester, TestGrid's AI teammate that takes plain-English intent and turns it into runnable, maintainable test automation for web and mobile applications.
- Designed and shipped the agent platform as a secure, multi-tenant service: conversational memory, per-customer knowledge bases, and behaviour tailored per client and per application, exposed through clean APIs the product consumes.
- Built the natural-language-to-execution pipeline end to end: interpret the intent, understand the application under test, resolve it to the right on-screen targets, and emit validated, platform-ready steps.
- Fine-tuned the underlying models for structured, directly executable output, and extended them to multi-step workflows and long sequences of UI interactions.
- Built dedicated element parsers for web, Android and iOS, and optimised how page data is prepared for the model, cutting noise and making generated tests measurably more reliable.
- Added an agentic element layer so elements curated for a specific page take priority during generation, producing tests aligned to what matters on that screen.
- Integrated vision-language understanding and inline image input, plus Figma knowledge-base upload, so tests can be generated from a design before a build exists.
- Engineered self-validation and auto-correction so generated steps are verified and repaired before execution, sharply reducing broken and flaky output reaching customers.
- Hardened it for production: authentication, request-level security controls, encrypted and versioned agent configuration, and staged rollout of new versions to live customers.
- Instrumented every interaction with usage, latency and cost telemetry, and used it to tune model selection and agent strategy for better answers at a lower cost per conversation.
- Owned delivery end to end, from packaging and release automation to production servers, across nine major versions driven by customer feedback.
Shipped modules
- Manual test case generation: built and deployed the API service behind it and wired it to the knowledge base, so full test cases generate from an uploaded document with no prompt required; extended it to multiple scenarios per case and a chat endpoint for on-demand generation.
- Page-aware keyword generation: reads the UI elements currently on screen and emits structured, executable web and mobile test cases aligned to the supported keyword set.
- Adaptive auto-healing: analyses a failing element and its metadata to find the root cause, then generates replacement elements and corrective steps on the fly.
Research and platform integrations
- Built an MCP server connecting Claude to TestGrid's device cloud, so a locally built web app can be launched and explored live in a cloud browser over secure tunnelling, with no manual deployment.
- Led the integration and technical evaluation of Deque Axe accessibility testing on cloud devices for web and mobile apps.
- Researched JMeter report extraction from command-line runs, improved the JSON dashboard metrics, and made the timeline report load summarised data far faster.
Machine learning and computer vision
- Integrated deep-learning image classification with Appium so tests identify UI elements visually and keep working when an app's internal properties change.
- Curated and labelled a custom dataset of more than a hundred mobile UI elements, retrained models with transfer learning, and applied YOLO object detection with non-maximum suppression, evaluated on mAP and IoU.
- Built an auto-healing layer that repairs affected tests when the UI changes, using a weighted longest-common-subsequence variant to re-anchor locators.
- Delivered a Python visual-regression engine using pixel comparison, template matching and OCR, with region-scoped comparison to remove false positives.
Python platform tooling
- Built a web element extraction framework with a lightweight Flask interface for managing and visualising extraction jobs, feeding clean structured data into AI test-case generation.
- Created reverse-tunnelling software that securely exposes locally hosted applications to remote infrastructure, with multi-server support for distributed execution against private environments.
- Architected a modular Selenium-based automation backend with cryptography, JavaScript execution and multiple-dispatch support, plus performance testing and network traffic analysis across browsers.
Beyond the code
- Onboarded and mentored a new engineer, assigned development work and reviewed their output.
- Reviewed blogs, documentation, slide decks and product pages with the design team for technical accuracy and brand consistency.