Extracting knowledge from unstructured or semi-structured textual information is essential for the machine learning applications that power DoorDash's search experience, and the development and maintenance of its product knowledge graph. Large language models (LLMs) have opened up new possibilities for utilizing their power in these areas, replacing or complementing traditional natural language processing methods. LLMs are also proving to be useful in the label and annotation generation process, which is critical for these use cases. In this talk, we will provide a high-level overview of how we incorporated LLMs for search relevance and product understanding use cases, as well as the key lessons learned and challenges faced during their practical implementation.

Traffic in optical backbone networks is increasing and becoming more heterogeneous with respect to bandwidth and QoS requirements due to the popularity of high-bandwidth services (such as cloud computing, e-science, telemedicine, etc.), which need to coexist with traditional services (HTTP, etc.). Mixed-line-rate (MLR) networks that support lightpaths of different rates such as 10, 40, 100 Gb/s, etc., are being studied to better support the heterogeneous traffic demands. Here, we study the important topic of degraded services in MLR networks, where a service can accept some degradation (i.e., reduction) in bandwidth in case of a failure in exchange for a lower cost, a concept called partial protection. Network operators may wish to support degraded services to optimize network resources and reduce cost. We propose using multipath routing to support degraded services in MLR networks, a problem that has not been studied before and is significantly more challenging than in single-line-rate (SLR) networks. We consider minimum-cost MLR network design (i.e., choosing which transponder rates to use at each node), considering the opportunity to exploit multipath routes to support degraded services. We propose a mixed-integer-linear-program (MILP) solution and a computationally efficient heuristic, and consider two partial-protection models. Our illustrative numerical results show that significant cost savings can be achieved due to partial protection versus full protection and is highly beneficial for network operators. We also note that multipath routing in MLR networks exploits volume discount of higher-line-rate transponders by cost-effectively grooming requests over appropriate line rates to maximize transponder reuse versus SLR.

As new Internet applications are emerging, traffic in optical backbone networks is increasing with heterogeneity in bandwidth and QoS requirements. Higher bit-rate wavelength channels of 40 Gbps, 100 Gbps, and beyond are being deployed in WDM backbone networks to meet these growing traffic needs. Mixed-line-rate (MLR) networks that support various line rates on the same fiber are becoming popular to address the heterogeneous traffic growth. We study the important problem of long-term cost-effective dimensioning of WDM backbone networks with MLR. Our study should enable operators to dimension their networks with deployment of MLR transponders and optical cross-connect (OXC) ports over multiple time periods. We study two multi-period dimensioning approaches: all-periods planning and incremental planning. Our approaches predict the network equipment to be deployed at various nodes in the network during each time period while exploiting the cost reduction of network equipment over time to support traffic growth over multiple time periods. We propose mixed-inter-linear-program (MILP) solutions and a computationally-efficient heuristic. Our illustrative examples show that significant cost savings can be achieved by our long-term planning approaches.

By utilizing Layer-1 Virtual Private Networks (L1VPN), a single physical network, e.g., optical backbone networks, can support multiple virtual networks, which is the basic infrastructure for cloud computing and other enterprise networks. The L1VPN hose model is an elegant and flexible way to specify the customers' bandwidth requirements, by defining the total incoming and outgoing demand for each endpoint. Furthermore, multi-domain physical infrastructures are common in L1VPNs, since these are usually deployed on a global scale. Thus, high-performance Routing for Multi-domain VPN Provisioning (RMVP) for the hose model is an important problem to efficiently support a global virtual infrastructure. In this paper, we formulate the RMVP problem as a Mixed Integer Linear Program (MILP). Also, we propose a Top-Down Routing (TDR) strategy to compute the optimal routing for the hose-model L1VPN in multi-domain backbone networks. Results indicate that TDR approaches the minimum routing cost when compared to ideal case of single-domain routing.

This paper proposes a Computational Intelligence framework for network security management. The framework uses reinforcement learning and Bayesian methods to achieve cross-layer optimization in heterogeneous, multi-layer wireless and wireline networks. Metrics based on the reputation of a node are used to measure performance. OPNET simulations results indicate that routing algorithms based on the newly proposed framework exhibit nearly four times improved network performance measured in terms of goodput.