Carlos Miguel Dimacali

Led a 7-engineer Gemini applied-AI workstream delivering production LLM capabilities for Google Cloud enterprise customers. Directed launch of Gemini-powered enterprise search driving $28M incremental ARR in the first year. Reduced serving cost per million tokens by 46% using speculative decoding, KV-cache quantization, and workload-aware batching on TPU v5e. Improved retrieval performance across three production layers by profiling React rendering, tracing Node.js API latency, and optimizing MongoDB aggregation pipelines. Introduced an evaluation framework combining expert rubrics with LLM-as-judge scoring, achieving 91% agreement with specialist reviewers. Mentored engineers to senior promotions and established repeatable promotion and design-review standards.

Led a 7-engineer Gemini applied-AI workstream delivering production LLM capabilities for Google Cloud enterprise customers. Directed launch of a Gemini-powered enterprise search capability generating $28M incremental ARR in the first year. Reduced serving cost per million tokens by 46% using speculative decoding, KV-cache quantization, and workload-aware batching on TPU v5e. Improved retrieval performance across three production layers by profiling frontend rendering, tracing Node.js API latency, and optimizing MongoDB aggregation pipelines. Introduced an evaluation framework using expert rubrics and LLM-as-judge scoring, achieving 91% agreement with specialist reviewers. Mentored engineers toward senior promotion and established repeatable promotion, architecture, and design-review standards.

Owned ranking and recommendation improvements for Facebook News Feed candidate generation across model development, distributed training, experimentation, and privacy review. Designed a two-tower PyTorch retrieval model improving candidate recall by 12.4% and driving a 3.1% sitewide lift in time spent during A/B testing. Scaled distributed training from 32 to 512 GPUs using FSDP and bfloat16, reducing full-refresh training time from 92 hours to 11 hours. Partnered with privacy engineering to deliver a differentially private fine-tuning workflow approved through Meta’s privacy review, unblocking four production launches.