Archana Kamath

Archana Kamath is an engineering leader whose work at DigitalOcean spans compute, networking, data protection and AI infrastructure. Her teams have developed cloud services including GPU offerings, backups and snapshots. In her 2026 AI Engineer appearance with Tyler Gillam, she was introduced as vice president of engineering for inference engine and AI infrastructure. Their shared argument for request-specific model routing applies familiar infrastructure concerns—cost, speed and service continuity—to the choice of models inside AI applications.

From networking to cloud infrastructure leadership

Kamath studied electronics and communications at Visvesvaraya Technological University in India and earned a master’s in electrical engineering at Penn State. Her study of network-on-chip ASIC design drew her toward networking as the field moved from physical switches and routers toward software-defined systems. She worked in Cisco Systems’ data-center group, developing a background in C programming and complex systems. She also led Cisco’s San Jose Women in Science and Engineering chapter and returned to Penn State to encourage women to pursue STEM careers. Her engineering and mentorship background connects technical practice with an interest in helping others enter and advance in the field.

At DigitalOcean, she progressed from engineering management in networking to broader responsibility for compute and networking products. A September 2025 profile described her as senior director of Compute and Network, leading teams responsible for Droplets, virtual private clouds, load balancers, cloud firewalls and IP-addressing capabilities. Her teams’ milestones included expanding GPU offerings with AMD MI300X accelerators and improving backup and snapshot performance. She emphasized making complex infrastructure straightforward for customers, shipping useful increments and using customer feedback to guide further development.

Her co-authored account of DigitalOcean’s backup and snapshot redesign explains what that approach required in practice. The team addressed both how much recent data customers could lose and how quickly they could recover after a failure. More efficient changed-block tracking reduced the data that needed processing; compression and flash storage improved the backup path; parallel image downloads accelerated restoration. Distributing backup start times reduced concentrated load, while a distributed queue replaced schedule polling. The improvements required coordinated work across storage, networking and scheduling, organized around customers’ recovery needs.

Choosing models for each request and planning for failover

Her 2026 work on inference and managed agent orchestration extended those operational concerns into AI. In Preferences Over Benchmarks: Model Routing, Kamath and Gillam argued that a public leaderboard cannot determine the right model for every request. Classification and labeling may work well with a small open model; inline code completion demands fast responses; accuracy-sensitive code review and security work may justify a frontier model. Selection also depends on the system prompts and tools surrounding the model, the budget, required latency and the user’s preferences. These examples describe a decision process rather than a universal ranking of models.

Their case for routing includes failover as a production requirement. An application dependent on one model needs another way to serve requests when that model becomes unavailable or its performance deteriorates. Gillam, who built parts of the router and led the demonstration, showed how developers could configure that behavior: a manually ranked model pool favored one model and used another if it failed, while a fastest-model policy selected from a pool using recent latency. Developers could define tasks in natural language, set preferences and layer decision rules onto presets. Those controls made routing choices explicit enough to test and revise.

Testing routing against cost and quality

Gillam demonstrated the approach in an OpenCode workflow that built a spinning-wheel application, added unit tests and wrote a README. Requests passed to models configured for different tasks, while an observability panel displayed model selection, token usage and accumulating cost. The presenters reported a total session cost of 14 cents for the router and 44 cents for the premium-model comparison. They judged the outputs broadly similar, but acknowledged that the visual comparison was subjective. A separate evaluation reported 90% correctness for the router and 95% for the comparison model. These were results from their demonstration and evaluation, not a general guarantee of equivalent quality or savings.

The joint presentation treated workload-specific evaluation as the means of improving those decisions: route requests, evaluate the results on the application’s own tasks, adjust the configuration and repeat. It also placed caching and personalization above routing. Caching avoids paying repeatedly for the same answer; personalization aims to adapt selection to what works for a team over time. Their comparison of model orchestration with cloud cost optimization framed inference spending as an ongoing engineering responsibility, alongside output quality, response time and availability.

Developing engineering talent

Kamath’s leadership interests also include how engineering organizations develop talent. In an interview about her path into technical leadership, she advocated mentorship, structured hiring and promotion, and greater visibility for women leaders. That emphasis complements her earlier outreach: advancement depends on how organizations recognize, support and develop engineers as well as on the systems those engineers build.