Highlights
- +89% engineering delivery output — merged pull requests across all repos, while freeing the team to focus on higher-value work
- −81% human-reported bugs — down to a full-year low, while output nearly doubled
- 2× code-review throughput — reviewers absorbed double the PR volume with zero added latency, including dedicated security review
- Claude Code across the entire lifecycle — 47 custom skills, 6 subagents, and 19 slash commands turning team conventions into governed, reusable workflows
- Precisely scoped releases — regression testing targets only what changed and what’s deployed, with release cuts computed and confirmed at hard gates
- Continuous, automated incident response — production errors are triaged and resolved on an ongoing basis, read-only and security-conscious, with the large majority of issues already closed
- 10× growth in AI operations in a single month, with a 92% autonomous-run success rate
Client
An iconic US landmark and one of the country’s most recognizable tourist attractions. SPD Technology serves as its primary software engineering partner and staff-augmentation engineering team.
Product
The Client’s guest-admission and venue ticketing platform, built on a single monorepo that consolidates the backend, four frontends, BrightSign digital signage, the automation-test suite, and infrastructure-as-code. SPD Technology delivers across this entire codebase, which handles ticket sales and admission for the attraction. On top of it, an operations layer connects the platform’s release process and production monitoring (Jira, Slack, Azure DevOps, Sentry) to guarded AI automation.
Goals and objectives
Development & quality
- Codify engineering conventions into reusable, enforceable workflows instead of leaving them as tribal knowledge.
- Keep code review, including payments/PII security, ahead of rising PR volume rather than trading quality for deadline pressure.
- Keep documentation, specs, and the knowledge base continuously in sync with code, so they stay trustworthy as the platform evolves.
- Accelerate and de-risk the translation of design files into front-end code.
Release & operations
- Free developer time from routine tickets for higher-value work.
- Scope regression testing precisely to what actually changed and what is actually deployed, instead of broad over-testing.
- Make release cuts computed and repeatable rather than manual and error-prone.
- Investigate production incidents quickly while minimizing security exposure, resolving the access-vs-exposure tension.
Overall: raise delivery throughput and quality without adding headcount.
Project challenge
As delivery accelerated, pain points across development and operations threatened quality, security, and velocity.
In daily development
- Conventions lived in people’s heads rather than being codified, so consistency and onboarding depended on tribal knowledge.
- Code review could not keep pace with PR volume, creating security risk under deadline pressure, especially around payments and PII.
- Documentation drifted out of sync with code, eroding trust in specs and the knowledge base.
- Design-to-UI translation was slow and imprecise, handled manually from design files to front-end code.
In release & operations
- Routine tickets consumed developer time that could have gone to higher-value work.
- Teams over-tested because regression scope was unclear relative to what was actually deployed.
- Release-cut decisions were manual and error-prone, relying on cross-referencing Jira and git history by hand.
- Incident debugging pulled against security: fast resolution needed broad access, which conflicted with minimizing exposure.
Solution
SPD Technology embedded Claude Code into the engineering workflow as an assistive layer spanning development, quality assurance, releases, and production operations.
AI-assisted development and engineering standards
SPD codified the team’s established engineering practices into 47 custom skills, 6 subagents, and 19 slash commands, turning conventions that previously depended on individual knowledge into reusable workflows.
Feature and bug-fix workflows use phased review gates and stop before PR creation, ensuring engineers remain responsible for reviewing and approving changes.
Code and security review
Claude-assisted review provides an additional quality layer before human review.
Dedicated workflows support code review, security checks, and re-evaluation of open PR discussions against the latest version of the code.
This allows AI to help engineers identify potential issues earlier while human reviewers retain final responsibility for approving changes.
Living documentation
Ten documentation-audit skills help keep CLAUDE.md, specifications, and the project knowledge base aligned with the codebase.
Automated verification identifies potential discrepancies between documentation and implementation so engineers can address them as the platform evolves.
AI-assisted testing
Playwright and AQA authoring workflows help engineers create and maintain automated tests.
Additional test-fidelity controls are designed to identify weakened assertions, skipped checks, and other situations where an AI-generated test could appear successful without adequately validating functionality.
Release and regression management
AI-assisted regression analysis compares release changes with what is actually deployed, maps those changes to affected business modules, considers feature-flag state, and generates a prioritized regression scope.
Release-cut automation similarly brings together ticket and source-control information to identify the appropriate release commit and potential blockers.
Remote-facing actions remain protected by explicit human confirmation gates.
Production incident investigation
Read-only production investigation tooling helps analyze production errors and turn previously unstructured monitoring signals into actionable engineering work.
The approach gives engineers faster access to diagnostic information without giving AI unrestricted authority over production systems.
AI operations orchestration
SPD has also developed a Jira- and Slack-triggered AI Agent Orchestrator designed to assist with routine engineering and operational workflows.
Guardrails prevent unsupervised production actions, keeping engineers in control while allowing repetitive work to be handled more efficiently.
Regression-scope analysis, release-cut automation and incident investigation are live; the AI Agent Orchestrator is in final rollout.
Technologies Used: Claude Code (custom skills, subagents, slash commands); cloud CI/CD pipelines with PR validation and branch policies; automated end-to-end testing; issue-tracking and team-messaging integrations; production monitoring; CLAUDE.md, behaviour specifications, and a project knowledge base; dependency graph, feature matrix, feature flags; dedicated AI-agent infrastructure.
Results
Delivery & quality (2026 vs. a flat 2025 baseline, at constant headcount)
- +89% engineering delivery output (merged pull requests across all repos)
- 2× code-review throughput (reviewers absorbed double the PR volume with zero added latency)
- −81% human-reported bugs (down to a full-year low while output nearly doubled)
- 6 platform-scale initiatives shipped concurrently (monorepo, CI, IaC, test suite, BDD specs, full documentation)
Sustaining +89% merged output without AI would have required roughly 0.9× additional engineering capacity. The team effectively doubled output at the same size.
Operational performance (measured monthly volumes, August 2026)
- 10× AI operations scale in a single month (from 38 pipeline runs in July to 443 in August)
- 92% autonomous-run success rate at production volume
- ~200 production errors auto-triaged and resolved per month (converted from untriaged Sentry noise; 77% already closed)
- 437 production incidents auto-investigated (a net-new capability that did not exist as tracked work before July)
Incident investigation is now a continuous automated stream rather than an on-demand manual task: issues surface and close continuously instead of waiting to be noticed. Releases are more precisely scoped, with regression testing targeting what actually changed and what is actually deployed, and each release cut from a computed, minimal-drift commit with hard confirmation gates. Routine triage and release mechanics run as guarded automation, returning developer time to higher-value work, while read-only investigation and per-step confirmation gates keep security exposure low.