github·Mar 12, 2026·3 min readCurated summary

Continuous AI for accessibility: How GitHub transforms feedback into inclusion

GitHub built a continuous, AI-assisted accessibility feedback system to replace scattered reports, unclear ownership, and unresolved “phase two” promises. The workflow combines GitHub Actions, Copilot, and GitHub Models to turn user feedback into tracked, prioritized issues while preserving human judgment. Its goal is continuous follow-through: every accessibility barrier is captured, routed, reviewed, and acted upon. ## Accessibility as a Living System - GitHub treats accessibility as an ongoing methodology rather than a one-time audit or standalone product. - The approach combines: - Automation - Artificial intelligence - Human expertise - Real user feedback is considered more valuable than automated code scans because it reveals barriers experienced in real workflows. - The system supports GitHub’s 2025 Global Accessibility Awareness Day pledge to improve accessibility across the open source ecosystem. - Technology helps process feedback at scale, turning unstructured reports into clearer, implementation-ready work. ## Designing for Different Users The workflow was designed around three primary groups: - **Issue submitters** - Community managers, support agents, and sales representatives submit reports for users and customers. - Since they may not be accessibility specialists, the system guides them and teaches accessibility concepts during submission. - **Accessibility and service teams** - Engineers and designers need actionable reports containing reproducible steps, WCAG references, severity ratings, and ownership information. - **Program and product managers** - Leaders need trend data, issue categories, and progress visibility to prioritize investments. The design treats feedback as data moving through a pipeline and allows the process to evolve over time. ## Event-Driven Feedback Workflow - Each workflow stage triggers a GitHub Action that determines what happens next. - Key events include: - New issues launching Copilot analysis through the GitHub Models API - Status changes initiating hand-offs between teams - Resolutions triggering follow-up with the original submitter - Actions can be started manually or rerun, allowing humans to intervene whenever necessary. - GitHub initially built the system largely by hand in mid-2024; newer tools such as Agentic Workflows could now create similar Actions from natural-language instructions. - The workflow contains seven stages: - Intake - Copilot analysis - Submitter review - Accessibility team review - Link audits - Closing the loop - Improvement - Feedback loops allow submitters to rerun analysis, resolved issues to return for further review, and improvements to update Copilot prompts. ## Actioning Intake - Accessibility feedback can arrive through support tickets, social media, email, direct outreach, or GitHub’s accessibility discussion board. - Approximately 90% of feedback currently comes through the public discussion board. - Public discussions let other users: - Confirm reported problems - Add context - Share workarounds - Reports from the community often contain more detail than conventional support tickets. - GitHub acknowledges every report within five business days, including reports it cannot directly address. - When internal action is needed, a team member creates a tracking issue using a custom accessibility feedback template. - The template records: - The user’s original report - The feedback source - Relevant product components - This preserves important context as feedback moves from intake into triage.

Read original(opens in new tab)

Event Driven ArchitectureGithub CopilotGitHub ActionsAgentic Workflows+3

naver·Dec 3, 2025Original article

Naver TV (opens in new tab)

Naver’s VLOps framework introduces an event-driven approach to MLOps, designed to overcome the rigidity of traditional pipeline-based systems like Kubeflow. By shifting from a monolithic pipeline structure to a system governed by autonomous sensors and typed messages, Naver has achieved a highly decoupled and scalable environment for multimodal AI development. This architecture allows for seamless functional expansion and cross-cloud compatibility, ultimately simplifying the transition from model training to large-scale evaluation and deployment. ### Event-Driven MLOps Architecture * Operations such as training, evaluation, and deployment are defined as "Typed Messages," which serve as the primary units of communication within the system. * An "Event Sensor" acts as the core logic hub, autonomously detecting these messages and triggering the corresponding tasks without requiring a predefined, end-to-end pipeline. * The system eliminates the need for complex version management of entire pipelines, as new features can be integrated simply by adding new message types. * This approach ensures loose coupling between evaluation and deployment systems, facilitating easier maintenance and infrastructure flexibility. ### Omni-Evaluator and Unified Benchmarking * The Omni-Evaluator serves as a centralized platform that integrates various evaluation engines and benchmarks into a single workflow. * It supports real-time monitoring of model performance, allowing researchers to track progress during the training and validation phases. * The system is designed specifically to handle the complexities of Multimodal LLMs, providing a standardized environment for diverse testing scenarios. * User-driven triggers are supported, enabling developers to initiate specific evaluation cycles manually when necessary. ### VLOps Dashboard and User Experience * The VLOps Dashboard acts as a central hub where users can manage the entire ML lifecycle without needing deep knowledge of the underlying orchestration logic. * Users can trigger complex pipelines simply by issuing a message, abstracting the technical difficulties of cloud infrastructure. * The dashboard provides a visual interface for monitoring events, message flows, and evaluation results, improving overall transparency for data scientists and researchers. For organizations managing large-scale multimodal models, moving toward an event-driven architecture is highly recommended. This model reduces the overhead of maintaining rigid pipelines and allows engineering teams to focus on model quality rather than infrastructure orchestration.

Event Driven ArchitectureMultimodal AIOrchestrationMlops+4

toss·Nov 30, 2025Original article

From Legacy Payment Ledger to Scalable System (opens in new tab)

Toss Payments successfully modernized a 20-year-old legacy payment ledger by transitioning to a decoupled, MySQL-based architecture designed for high scalability and consistency. By implementing strategies like INSERT-only immutability and event-driven domain isolation, they overcame structural limitations such as the inability to handle split payments. Ultimately, the project demonstrates that robust system design must be paired with resilient operational recovery mechanisms to manage the complexities of large-scale financial migrations. ### Legacy Ledger Challenges * **Inconsistent Schemas:** Different payment methods used entirely different table structures; for instance, a table named `REFUND` unexpectedly contained only account transfer data rather than all refund types. * **Domain Coupling:** Multiple domains (settlement, accounting, and payments) shared the same tables and columns, meaning a single schema change required impact analysis across several teams. * **Structural Limits:** A rigid 1:1 relationship between a payment and its method prevented the implementation of modern features like split payments or "Dutch pay" models. ### New Ledger Architecture * **Data Immutability:** The system shifted from updating existing rows to an **INSERT-only** principle, ensuring a reliable audit trail and preventing database deadlocks. * **Event-Driven Decoupling:** Instead of direct database access, the system uses Kafka to publish payment events, allowing independent domains to consume data without tight coupling. * **Payment-Approval Separation:** By separating the "Payment" (the transaction intent) from the "Approval" (the specific financial method), the system now supports multiple payment methods per transaction. ### Safe Migration and Data Integrity * **Asynchronous Mirroring:** To maintain zero downtime, data was initially written to the legacy system and then asynchronously loaded into the new MySQL ledger. * **Resource Tuning:** Developers used dedicated migration servers within the same AWS Availability Zone to minimize latency and implemented **Bulk Inserts** to handle hundreds of millions of rows efficiently. * **Verification Batches:** A separate batch process ran every five minutes against a Read-Only (RO) database to identify and correct any data gaps caused by asynchronous processing failures. ### Operational Resilience and Incident Response * **Query Optimization:** During a load spike, the MySQL optimizer chose "Full Scans" over indexes; the team resolved this by implementing SQL hints and utilizing a 5-version Docker image history for rapid rollbacks. * **Network Cancellation:** To handle timeouts between Toss and external card issuers, the system uses specific logic to automatically send cancellation requests and synchronize states. * **Timeout Standardization:** Discrepancies between microservices were resolved by calculating the maximum processing time of approval servers and aligning all upstream timeout settings to prevent merchant response mismatches. * **Reliable Event Delivery:** While using the **Outbox pattern** for events, the team added log-based recovery (Elasticsearch and local disk) and idempotency keys in event headers to handle both missing and duplicate messages. For organizations tackling significant technical debt, this transition highlights that initial design is only half the battle. True system reliability comes from building "self-healing" structures—such as automated correction batches and standardized timeout chains—that can survive the unpredictable nature of live production environments.

Event Driven ArchitectureDatabase DesignApache KafkaMicroservices+4