▶ Watch ↗AI Engineer Code 202518:08
Conference affiliation: Railway · 2025
Mahmoud Abdelwahab is a senior developer relations engineer at Railway who builds and explains tools for AI-assisted software development and operations. His projects give agents concrete database and infrastructure context, preserve coding environments between sessions, and make proposed repairs available for human review. His writing also examines where these tools fall short: an agent can produce a functioning application while missing everyday product behaviors, or reach a destructive API that bypasses safeguards elsewhere in the system.
Mahmoud’s earlier work taught application development through complete projects. His Prisma teaching series developed a full-stack application across data modeling, a GraphQL API, authentication, image uploads, and deployment. His exploration of database access from React Server Components adapted an early React demonstration to Prisma, showing how database queries could sit within server-rendered components while retaining type-safe access.
His writing for WorkOS addressed enterprise authentication. A multifactor authentication guide walked developers through adding time-based one-time passwords and SMS verification. Across these projects, Mahmoud followed features through the application, explaining how the pieces fit together beyond an isolated API call.
By 2023, Mahmoud was a developer advocate at Neon, working with PostgreSQL, browser applications, and language-model tools. He helped build and documented the Postgres AI Playground, an experiment originating in an internal hackathon. It connected a browser to Postgres through Neon’s serverless driver and supplied the database’s table and column definitions alongside the user’s request to a language model. That schema-aware SQL generation gave the assistant concrete information about the database it was being asked to query or change.
Another collaborative Neon project, the YC idea matcher, converted a startup description into an embedding and retrieved similar company descriptions stored in Postgres. Mahmoud explained how differently worded requests could express the same idea: matching depended on their vector representations rather than shared keywords. The project also demonstrated how vector search could fit into an existing relational database without a separate store. His explanation of SQL template tags addressed a complementary concern: separating query text from parameters to prevent SQL injection.
At Railway, Mahmoud’s work includes persistent coding environments and tools for operating deployed services. His Agents Anywhere project packages Claude Code into an always-on Railway container accessible over SSH; ChatGPT Codex Anywhere provides a persistent remote Codex environment. Persistent storage keeps repositories and agent configuration available across redeploys, while SSH lets users reconnect to the same workspace from different devices.
His railway-autofix prototype connects Railway observability to a coding agent that investigates incidents and opens inspectable GitHub pull requests. It brings together service architecture, deployment logs, resource consumption, HTTP errors, and latency so the agent can investigate beyond a single alarm. Slow database queries, for example, can affect an application even when CPU and memory appear healthy. His AI Engineer Code demonstration uses Inngest durable workflows and a Railway-hosted OpenCode server to resume failed operations without repeating completed steps. The proposed repair remains available for review before it is merged.
Mahmoud’s infrastructure writing also explains the limits of restricting an agent’s tools. In his 2025 account of Railway’s MCP server, he described the team’s choice to reuse the Railway CLI for authentication and operations. Reuse reduced duplicated machinery while exposing weaknesses in the CLI’s programmatic experience. Omitting deletion tools removed one route to destructive behavior, but an agent able to execute arbitrary CLI commands could still perform dangerous operations.
A subsequent production-database deletion incident illustrated that gap. An agent used an account-scoped token to call a legacy API endpoint whose deletion behavior lacked the dashboard’s recovery window. Railway recovered the customer’s data and changed the API to delay deletion for 48 hours. Mahmoud argued for recoverable infrastructure operations, credentials scoped to the task, and protections that apply through the interfaces agents actually use. The recovery and API changes were Railway team work; his account explained the failure and its implications for infrastructure design.
Mahmoud’s application-building work examines what remains after an agent gets a feature running:
▶ Watch ↗AI Engineer Code 202518:08
▶ Watch ↗AI Engineer NYC 2026 Online Talks13:21
Loading more talks and highlights…