PubMatic, the publicly traded programmatic advertising platform processing nearly 350 trillion ad impressions per year, deployed AI Architect across 450 engineers and 500+ repositories and cut the planning overhead that consumed the first two weeks of every large project. Technical design time on large epics dropped from approximately 2 weeks to 3 days. Over an 8-week evaluation, the team shipped twice the volume of work on repositories where AI Architect was active.
2x
Volume of work shipped on indexed repos over 8 weeks
+44% net
Lift in merge volume vs. non-indexed repos
2 weeks to 3 days
Technical design time on large epics
2 weeks to 1-2 days
Time for ticket-to-PR on mid-sized tasks
500+
Repos indexed
450
Engineers
The Challenge
Coding agents have compressed implementation timelines dramatically, but technical design and planning have remained largely manual. Before anyone writes a line of code, someone still has to trace which services are affected, understand how they connect, and gather the institutional context that lives in people’s heads rather than in documentation.
PubMatic’s platform processes around 840 billion daily ad impressions across infrastructure that has been evolving for over 15 years, spanning more than a thousand repositories and hundreds of services. When a large epic came in, the first two weeks were spent tracing affected services, reading unfamiliar code, scheduling meetings with other teams, and chasing down decisions that had never been written down.
Mid-sized tasks followed the same pattern. The bottleneck was never typing speed. It was always the time it took to understand what to build and where it fits.
“As the organization grows, complexity grows with it. We were actively looking for ways to bring AI into the design process, because every large epic started the same way, a senior engineer trying to reconstruct how a dozen services fit together, then waiting on other teams to confirm what they already knew but had never documented.”
AI Architect indexed PubMatic’s full engineering surface into a single knowledge graph: more than 500 repositories, years of commit history, thousands of Jira tickets, and thousands of Confluence documents. Connected first to Jira and then to Slack, AI Architect became available to every engineer, PM, and tech lead across the organization.
System-level knowledge graph: Indexed code, commits, Jira tickets, Confluence docs, and PR comments into a single queryable graph spanning 500+ repos.
Automatic feasibility analysis in Jira: Tagging @bito on an epic produced a technical plan grounded in the real system, with adoption spreading organically across teams.
Conversational access in Slack: Any engineer could pull AI Architect into a conversation to size an epic, evaluate a design, or check downstream impact.
Scalability and pattern-aware planning: Flagged recurring scalability concerns from incident post-mortems and sequenced parallel workstreams across a six-service, three-language epic, matching work to the right teams.
Institutional knowledge in every plan: On a UI epic, AI Architect recognized from planning docs and recent commits that PubMatic was running a company-wide i18n initiative, and generated the plan with externalized strings matching the emerging patterns. Without AI Architect, that kind of cross-initiative awareness surfaces only during review or after merge as rework.
“Things like catching our i18n patterns on its own, that is what is really different. AI Architect is reading the same signals a senior engineer would, our tickets, our docs, our review comments, and planning the way someone who knows us would plan.”
Technical design time on large epics dropped from approximately 2 weeks to 3 days: Engineers went from ambiguous epic to reviewable technical plan in days, eliminating most of the cross-team meetings that had anchored design timelines.
44% net lift in merge volume over 8 weeks: Indexed repos grew merges 48% during the evaluation window. Non-indexed repos with similar team sizes, repo sizes, and baseline activity grew 4% over the same window. The control confirms the uplift came from AI Architect.
Twice the volume of work shipped on indexed repos: No increase in reverts or post-merge incidents observed during the evaluation period. The velocity gains did not come at the cost of quality.
Standout Result
Controlled 8-week evaluation confirmed velocity doubled without quality tradeoff: PubMatic indexed a subset of repositories and kept comparable repos with similar team sizes and baseline activity as a control group. Indexed repos shipped twice the volume of work and 44% more merges over the evaluation window. Control repos stayed flat. Revert rates and post-merge incidents did not rise. The gains were sustained across the full 8 weeks.
“Once AI Architect moved into Slack, adoption grew through the organization fast. Anyone could pull it into a conversation, ask it to evaluate an idea, or have it draft a plan. It stopped being something the architect ran and became something the whole team worked with.”
Every large engineering organization recognizes this pattern: the real cost of a feature is rarely in the code. It is in the weeks of context-gathering, meetings, and tribal knowledge transfer that have to happen before the code can be written. AI Architect attacks that cost directly by making system-level understanding available to every engineer and every coding agent, grounded in code, commits, Jira tickets, Confluence docs, and PR comments. At PubMatic’s scale, that shift compressed the entire cycle from epic to merge and freed senior engineers to do senior-engineer work.
About PubMatic
PubMatic (pubmatic.com) is the leading AI-powered ad tech company delivering digital advertising performance through an intelligent, unified platform. The company is publicly traded (NASDAQ: PUBM), has delivered 36 consecutive quarters of adjusted EBITDA profitability, and processed nearly 350 trillion ad impressions in 2025. PubMatic runs its own infrastructure and has been building its platform for over 15 years, supported by an engineering organization of approximately 450 engineers.
