Bringing AI-native engineering to a food delivery platform

Client overview

The client runs one of the largest food delivery platforms in North America, operating at a scale that few engineering teams experience. Their environment includes hundreds of microservices, cross-functional teams, and millions of daily user interactions. Coordinating these elements without disrupting users or breaking connected services requires process discipline at every level.

Coherent Solutions has been the client’s trusted engineering partner for over six years. The organization’s teams have the specific context required to work in the client’s complex environment, shaping the work and how it is completed.

The challenge

In a food delivery environment, the stakes are high. A payment failure or a data mismatch during peak hours isn’t just a simple service ticket — it’s lost revenue.

The client’s review queues were piling up as the organization’s senior engineers absorbed the coordination cost. Their team often reconstructed the investigation context from scratch across tickets, PRs, and incident threads, while fielding questions from other teams who depended on their answers. Although the system was producing more, delivery quality and timelines were less predictable.

The company was already using AI. Tools were in use across engineering teams, and individual engineers were moving faster on isolated tasks. Overall delivery wasn’t improving, however, and the reason wasn’t hard to find.

There is often a pattern among organizations at the AI Adoption stage: activity goes up, but outcomes remain unchanged. More tools don’t fix the issue. What’s missing is a connected delivery system, one where AI-supported work feeds a shared feedback loop at every stage, and human judgment supports important decisions. Until coordination is addressed in AI-driven systems, production speed gains just create more pressure in existing bottlenecks.

Our approach

Instead of layering on additional tools or roles, Coherent’s team identified where coordination was breaking down and rebuilt those connections. Working sprint by sprint, using each cycle's output to improve the next, the team built a connected delivery system with agents and capabilities spanning every major phase of delivery. They also managed the new system by creating a versioned AI Playbook, one that could be updated and tracked like production code.

Coherent deliberately began by addressing zero-risk tasks (documentation, diagrams, ticket creation) before touching any code. On the client’s high-load production system, one bad AI-assisted change early in the engagement would have ended the initiative. Building confidence in the work before touching the system was important for creating trust and establishing effective work processes.

The engagement covered five areas of focus:

• Investigation and discovery: MCP-connected agents automatically gathered data from multiple tools and posted structured diagnostic reports to Slack — turning hours-long investigations into minutes and giving on-call engineers a head start on root causes before issues occurred.

• Implementation and code review: A layered AI review system checked generated code for security, architecture, syntax, test coverage, and API design before escalating to human review — freeing engineers to focus on high-level judgments about system impact.

• Pull request (PR) automation: AI agents populated structured PR descriptions automatically. A local agent addressed reviewer comments, provided justifications, and resolved threads, while final approval remained with a human.

• Incident response: A Slack app listened for alerts and queried runbooks, logs, analytics, and Jira history — using the data to post a diagnostic report in the alert thread. One standing rule applied: only an engineer could make decisions affecting production data.

• Engineer onboarding: A personalized walkthrough, dynamically built from the codebase, documentation, and Slack history, guided new engineers joining complex service teams. Triggered by running a single command, the walkthrough was validated with positive engineer feedback before it became part of standard onboarding for teams embedded in the system.

What made these distinct developments into a cohesive delivery system was the feedback loop running beneath them all. During each sprint, every delivery phase produced data showing where agents performed well and where they needed improvement. That data went into the next round of AI Playbook updates, which were used to improve subsequent sprints and other projects. The tools didn't just optimize their respective cycles; they improved the entire system.

The solution

In three months, the team’s code generation quality was reliable enough for production use. In six months, the full system was running in every sprint as standard practice. During the process, the team maintained and versioned the AI Playbook, updating instructions, workflow patterns, tool integrations, and governance rules to ensure that each cycle built on the last rather than starting over.

One structural change proved to be pivotal on the adoption side. When AI contribution was voluntary, usage hovered around 5–10%. After the client set an 80% AI contribution target, engineers who had previously avoided AI began using it and quickly saw the benefits. When cultural encouragement couldn’t improve adoption rates, a requirement could. Throughout the entire engagement, this was the most transferable lesson: voluntary adoption usually produces pilot-level results, while a formal delivery target produces system-level change.

The client also designated an internal AI lead to build, test, and share automations, and asked on-call engineers to improve team processes using AI during their rotations. Those artifacts were shared across the organization and adopted by other teams, turning one team’s work into a reference point for the broader engineering organization.

The impact

Using Jellyfish to compare the new system’s cycle times across multiple sprints, the client found that times for complex work (large refactors, architectural changes, and comprehensive investigations) dropped 50–80% compared to the previous system, which used unstructured AI. The team found that an investigation that took fifteen hours to complete in the old system was completed in one hour in the new system. This specific incident’s improved time is an accurate representation of what the system change felt like for engineers doing the work.

When comparing the new and old systems, the difference isn’t AI tools — the company was already using them when the project began. The difference is that the new system is built around those tools, requiring human contribution, greater governance, and a feedback loop that improves processes and work with every cycle.

What’s next

This engagement shows what AI-native delivery looks like in production, not as a pilot, but running at scale for an organization with a large, complex engineering team, like the client’s. This project used Coherent’s Continuous Delivery Loop (CDL) framework, an approach that connects AI-powered delivery workflows with feedback loops, built-in governance, human oversight, and institutional knowledge that compounds with every cycle.

For organizations already using AI tools but not seeing consistent results across the delivery system, CDL is the right place to start.