See how Coherent Solutions helped a leading North American food delivery platform reduce complex engineering cycle times by 50–80% with AI-native delivery.
The client runs one of the largest food delivery marketplaces in North America, operating hundreds of microservices, petabyte-scale data pipelines, and millions of daily transactions on AWS. Coherent's team is embedded in the client's payments domain, working on two payment processors: a supervisor framework built on SQS and Cassandra, and a settlement process with a roughly 22-hour delay. Working within this system leaves little room for mistakes, which can lead to costly errors in customer payments.
At a glance
Client: A leading North American food delivery platform.
Industry: Food Tech & Food Delivery.
Challenge: Using AI to automate delivery in a high-load, payments system slowed by manual planning, review, and operations overhead. The client needed to safely integrate AI to avoid additional financial and customer risks .
Outcome: Ten stages of the delivery lifecycle automated end to end, with 100% of the planned rollout shipped in under six months, and AI-driven automation successfully handling live incidents.
Coherent's role: Embedded engineering partner, building the automation layer, the specialist agents, and the review and operations tooling running on top of it.
The challenge
Payments engineering teams carry a specific kind of overhead. Tickets require careful scoping, pull requests need security and performance reviews, and the context behind incidents must be reconstructed from logs, Slack threads, and data warehouses before they can even start diagnosing the problem. In addition to the built-in operational work, teams are actively working to ship new features.
The client wanted to use AI to absorb more of that overhead, freeing engineering teams to focus on refining and releasing new features. Integrating a technology like AI into payment systems, however, requires additional safeguards to avoid major risks for companies and their clients. Potential issues such as a hallucinated ID, a query against a partition-less table, or an unreviewed production webhook send could cause significant financial harm. For the client, the concern wasn't whether AI could write code for a live payments system, but rather it could be trusted to make judgment calls, and where that trust needs human validation.
Our approach
Coherent's team treated the automation layer as a codebase rather than a set of prompts. Everything the team built lived in a versioned, code-reviewed directory checked directly into the repository and structured around three parts.
-
Skills: Sixteen invocable workflows, one for each recurring task, from writing unit tests and drafting pull request descriptions to generating commit messages, building Jira tickets, and investigating a payment.
-
Agents: Five specialist sub-agents, each scoped to a narrow set of tools and a specific job, spawned by the skills that need them.
-
A knowledge corpus: Table schemas, enumeration values, and business-flow documentation, read into every session, so agents work from cached, verified context instead of rediscovering the same facts on every run.
The new automation layer covered ten stages of the team’s delivery lifecycle, including discovery, operations, and even code generation.
-
Discovery: One skill traced a data table's schema straight from source code, cross-referencing the underlying event class, its enums, and its Avro schema, then wrote the result into the knowledge corpus for every other skill to read. The corpus documents itself.
In planning, a Jira skill handled full ticket creation with exact custom-field mappings baked in, so story points, sprint assignment, and epic links landed correctly every time instead of depending on the model to remember conventions. Nothing got written until the engineer had confirmed the epic, the description, the estimate, and the sprint, in that order.
In code review, the team built its highest-leverage automation. A deep review spawns three specialist agents in parallel: one focused on security (OWASP checks, entitlement and brand isolation, webhook signature validation), one on performance (missing timeouts, unfiltered queries against partitioned tables, consumer lag), and one on test coverage. Two run on a more capable model tuned for judgment calls; the third runs on a lighter model tuned for pattern matching, because classifying test gaps doesn't need the same reasoning budget as flagging a security hole. All three report back on a fixed severity scale, with file, line, and fix.
In operations, an incident-investigation skill takes a UUID or a cart ID and fans out in parallel: a query against the data warehouse, a search across the team's incident and alert Slack channels, and a synthesis of the resulting payment timeline into one report. The team's most complex skill recovers failed partner-webhook events from a dead-letter queue, a job with real financial exposure if it goes wrong. It draws a hard line. Scripts own every piece of raw data handling, and the model only makes decisions and confirms them with a person, one webhook at a time, never in a batch.
One rule carried across every skill the team built: the model orchestrates and decides, but deterministic code touches the data. IDs, SQL, and exact field mappings live in scripts, not model output, because that's where hallucination and inconsistency do the most damage.
The impact
Nine months after the operations tooling went live, the AI-driven automation layer has fundamentally changed how the engineering team works. For example, the skill for dead-letter-queue recovery has generated dozens of incident folders, brand correction reports, and recovery runs XX. The client’s engineering team has also seen other system improvements:
-
Delivery cycle time from ticket to pull request dropped materially.
-
Code review overhead per pull request decreased by hours with the new parallel specialist agents in place.
-
Manual documentation effort on architecture and diagrams dropped to nearly zero.
-
Cross-team incident response speed increased by 10% to 50%, depending on the investigation.
None of these results came from giving the AI models maximum autonomy. Instead, Coherent focused on giving the agents narrower, well-defined jobs and adding human validation where the risk was highest.
What's next
Coherent will continue working with the client as an engineering partner. The dead-letter-queue recovery skill and the rest of the operations tooling continue to handle live incidents, and the payments team can extend the automation layer as new needs arise. What started as one engineer's build has since spread to other teams across the client's engineering organization through the shared internal repository, the same pattern the client has followed with Coherent for six years now, turning one team's tooling into a reference point for the next.
This engagement is a working example of Coherent's Continuous Delivery Loop (CDL) approach in action. CDL is the model Coherent uses in place of a traditional, linear SDLC; it’s built around continuous feedback loops instead of sequential handoffs, bridging the value gap in AI-assisted development. Learn more about CDL.
