Lin Qiao

UC Santa Barbara records Qiao’s computer-science doctorate in 2006. Its alumni award biography describes her subsequent research work at IBM’s Almaden Research Lab and experience with database, streaming, warehouse, logging, and metrics systems. Her TWIML guest biography also lists computer-science bachelor’s and master’s degrees from Fudan University, without graduation dates. These records establish a systems background before her later leadership in machine-learning infrastructure.[2][3]

Qiao’s UCSB biography identifies her as a senior director of engineering at Meta, leading PyTorch, AI compilers, and on-device AI platforms. A compiler translates a program into work that hardware can execute; the surrounding platform must also handle deployment and operational constraints. Her 2025 MAD Podcast interview describes leaving Meta in 2022 to start Fireworks. TWIML’s guest biography dates the Fireworks CEO and co-founder role to October 2022. Fireworks’ current founding-team page confirms that she still holds that position. Her role was engineering leadership within a broader team, rather than sole authorship of PyTorch.[2][4][3][1]

An accelerator has to fit the whole system

Qiao is one of the authors of the 2021 paper on Facebook’s first-generation inference accelerator deployment. Inference is the stage where a trained model processes new inputs. The paper describes how workload properties, including large models and substantial memory or network demands, shaped hardware and software together. It covers a platform and tooling ecosystem designed to support accelerators from different vendors, along with production deployment lessons.[5]

Where to start with her work

The MAD Podcast conversation is a useful first stop because it introduces PyTorch and Fireworks before moving into model deployment, function calling, hardware choices, and inference costs. Function calling concerns a model selecting structured actions for software to perform, which becomes relevant when an application must do more than return text. For an account closer to research engineering, follow it with the accelerator paper. The 2026 Index Ventures conversation adds Qiao’s account of the path through IBM, LinkedIn, and Meta that led to Fireworks.[4][5][6]

Qiao’s Fireworks work focuses on running and adapting open models for production applications. The company’s 2026 announcement describes a strategy of customizing models around businesses’ proprietary data and domain knowledge.[1][7][4]