A typical feature at the company passed through ten stages: requirements gathering, solution design, ticket creation, integration development, database work, backend code, testing, code review, deployment, and documentation. Those stages were not just process steps; they were different technology stacks. A single feature moved from API integration into SQL optimization, on into PHP, Python, or Scala, then into test frameworks and deployment infrastructure, and each crossing meant an engineer switching context or work waiting on the specialist who owned that layer. Requirements were often still being clarified while code was being written, forcing days of rework late in the cycle; documentation trailed deployment by weeks, so knowledge lived with individuals rather than the team. A standard feature took sixteen or more days to reach production.
The client’s growth trajectory made the arithmetic uncomfortable. Delivering more features under the existing model meant hiring proportionally more engineers, accepting longer cycles, or compromising quality, and none of those options fit a platform running tens of thousands of campaigns a day. The leadership question was sharper than developer productivity: could the company compress the entire lifecycle without weakening the quality gates that AdTech reliability depends on? Leadership chose to treat AI not as a coding accessory but as a structural answer to that question and engaged Zimetrics to engineer it
Zimetrics framed the engagement workflow-first and mapped with the client’s teams where delivery actually slowed: unclear requirements, undocumented decisions, manual test design, review queues, and repetitive analysis. The operating principle Zimetrics instituted now governs every AI-assisted task across the client’s engineering organization: Questions, then Options, then Agreement, then Implementation. Claude clarifies what is missing in a requirement, proposes two to three approaches with documented trade-offs, and only after engineers choose a direction does implementation begin. Engineers retain decision authority at every step; Claude strengthens the decisions rather than making them.
To keep assistance contextual rather than generic, usage is organized into four working modes:
• Developer Mode for brainstorming-first implementation
• SQL Mode for query development and optimization
• Code Review Mode for security and performance scanning
• General Mode for planning, research, and documentation
Rather than standardizing on a single model, Zimetrics matched Anthropic’s model portfolio to the work: Claude Haiku for lightweight tasks such as summarization and quick reviews, Claude Sonnet for the balance of speed and capability in code generation and PR review, Claude Opus for complex reasoning across architecture and multi-step debugging, and Claude Fable for domainspecific AdTech reasoning such as campaign logic. The engagement kept pace with the model frontier, adopting successive Opus releases from 4.1 through 5 as they shipped. All models run through the Anthropic API on Google Cloud Vertex AI, aligning with the client’s existing cloud estate and security posture. Claude was embedded where engineers already work, in Slack, GitHub, Jira, and Confluence, so adoption required no new workflow, only a better one.
The highest-return change came before any code existed. Zimetrics redesigned the front of the lifecycle so that Claude interrogates each incoming requirement for missing information, unvalidated assumptions, and edge cases, then decomposes agreed work into epics and Jira tickets complete with acceptance criteria, dependencies, and estimates. Design rework fell by half, and project setup became 60% faster as engineers started building with full context instead of discovering the requirement mid-implementation.
Zimetrics embedded Claude across PHP, Python, and Scala development, supporting implementation planning, code generation, and error-handling design, and generating the test suites that validate the work: PHP unit tests, Pytest integration tests, and Docker-based deployment simulation. Test coverage rose from 55% to above 70% while test development time halved. At review, Claude scans every pull request for security, performance, and reliability issues before a human reviewer opens it, turning a two-hour review into a focused fifteen-minute one. Twelve vulnerabilities were identified and fixed before reaching production, shifting security from a late-stage audit to an everyday engineering practice.
The clearest single win came from the client’s data platform. A recurring large-scale database update, spanning over 22 million rows across a year of campaign data, processed in 73 controlled batches, previously consumed four to six hours per batch. Zimetrics used Claude to analyze SQL execution plans, index strategy, partition pruning, and batch design, cutting it to 10 minutes batch, a 24-36X improvement that reduced compute cost and protected reporting SLAs. The same structured approach transformed data debugging: investigations that took two to four hours of manual analysis now reach root cause in 30 to 45 minutes, and query optimization reduced data scan costs by 35%
The operating model did not stop at engineering. Zimetrics’ project manager on the engagement used Claude to build their own delivery-management dashboard, consolidating Jira, GitHub, Slack, and leave data into live views of who is working on what, how workload is distributed, what is progressing or blocked, and what impact the work delivered. Leave planning is connected to actual workload, so an absence surfaces its dependent tasks and backup owner before it happens. In its first three quarters, the dashboard tracked 1,438 issues across 20 sprints, with 90% of the period’s tickets closed. A weekly-to-quarterly review rhythm, powered by Claude-generated summaries, shifted management conversations from counting tickets to measuring engineering outcomes. The same AI assistance that accelerated the code now runs the process around the code.
