Dmytro “Dima” Dzhulgakov is the co-founder and chief technology officer of Fireworks AI, which builds infrastructure for deploying and customizing open-weight generative models. A former PyTorch core maintainer and Facebook AI engineering lead, he approaches production AI through the constraints that determine whether applications actually work: latency, GPU capacity, reliability, and cost.
At Facebook AI, Dzhulgakov helped move deep-learning frameworks from research into production, working across PyTorch, Caffe2, model portability, and inference deployment. His early contributions addressed cross-platform inference and ONNX interoperability. In 2019, he co-authored research introducing the Deep Learning Recommendation Model, combining model-parallel embeddings with data-parallel computation for large-scale recommendation systems.
He co-founded Fireworks in 2022 with fellow AI infrastructure veterans. His open-source engineering projects include an adaptation of Llama inference code for Mistral and Mixtral models.
- Specialized open models: Dzhulgakov argues that domain-specific applications should use models adapted to their actual tasks instead of paying for unnecessary general-purpose capabilities. Smaller, fine-tuned models become particularly valuable when agents make repeated calls and latency accumulates.
- Inference economics under real latency constraints: He optimizes serving for application-specific response times, using custom CUDA kernels, workload-aware batching, and long-context prompt caching to increase throughput and reduce GPU costs.
- Shared infrastructure for customized models: Serving multiple LoRA adapters on shared GPU infrastructure makes specialized model variants practical without assigning dedicated hardware to every deployment.
- Tool-using compound AI systems: Fireworks’ FireFunction V2 supports conversational and parallel function calling. Dzhulgakov envisions applications that coordinate specialized models, retrieval, external APIs, structured outputs, and context-aware tool selection; his AI Engineer World’s Fair presentation illustrates that architecture through a multistep financial-research assistant.
