Olga Tsymboi, Dmitrii Stoianov, Ramil Latypov +11cs.CL
Data-residency constraints force enterprises to self-host LLMs, but continuous adoption of newer models without decommissioning their predecessors expands the serving fleet, fragmenting a finite GPU pool. We consolidate traffic from over 200 internal applications onto a single model by closing quality gaps identified through production error analysis along three axes: instruction following, function-calling, and internal task distribution. Quality is tracked by offline benchmarks stratified to production traffic and scored by deterministic verifiers or calibrated LLM judges. Rather than optimising all objectives jointly, which introduces cross-domain reward interference, we train a separate GRPO expert per axis and merge them via two-stage SLERP. Each expert's reward exposes a distinct failure mode, namely semantic collapse, over-calling, and verbosity hacking, each requiring a domain-specific fix. In non-reasoning mode the recipe surpasses a ${\sim}7\times$ larger by total parameters baseline on the in-house Arena with 69.6 to 65.8, instruction following with 0.85 to 0.83, and function-calling with 0.79 to 0.77, while lifting general dialogue benchmarks. The model absorbs 50% of platform traffic, 116M requests per month, at a fraction of the serving cost.
Micro-View Order-Dispatching assigns available drivers to passenger orders within each dispatch batch and is critical to the service quality and operational efficiency of ride-hailing platforms. Mainstream industrial solutions follow a multi-stage paradigm of model prediction, value calculation, and dispatch matching. Although dispatch quality is determined by the final batch-level assignment, these stages optimize different intermediate objectives. This cross-stage objective inconsistency means that improving a single stage does not necessarily improve the overall dispatch result. We therefore formulate Micro-View Order-Dispatching as a generative matching problem and propose GenMatch, an end-to-end Generative Matching framework and the first such framework deployed in a real-world production environment. Applying generative modeling to this problem introduces three challenges. First, each dispatch batch forms a dynamic sparse bipartite graph, requiring efficient structured batch-level encoding. Second, replacing the hand-crafted value function requires learning unified business utility from heterogeneous feedback. Third, directly generating an assignment requires tracking the evolving matching state because each selected order-driver pair changes the remaining feasible candidates. GenMatch addresses these challenges with a Context-Aware Bipartite Encoder, a Business-Aware Utility Learner, and a State-Aware Pointer Decoder. Extensive offline evaluations and online A/B tests in five cities across DiDi's international ride-hailing markets show consistent improvements over competitive baselines, confirming the effectiveness and practicality of GenMatch for industrial order-dispatching.
Modern e-commerce platforms often operate search, recommendation, personalization, and CRM systems independently, limiting opportunities for proactive customer re-engagement. This is particularly challenging for exploratory intents such as best smartphones or latest 5G phones, where users may leave the platform for external research before purchasing. We present a scalable, production-deployed framework that bridges search and CRM workflows through AI-powered Product Research Agents. The system identifies users with exploratory purchase intent and low engagement, conducts grounded multi-agent product research using behavioral signals, external knowledge, and enterprise catalog data, and delivers personalized recommendations through WhatsApp. We evaluate the framework in a 23-day production deployment involving approximately 15K WhatsApp notifications for mobile product discovery. The campaign achieved substantial CTR improvements over traditional WhatsApp recommendation campaigns, with evidence of secondary engagement through message forwarding and sharing. The deployment also generated downstream purchases and GMV impact, demonstrating the practical effectiveness of AI Product Research Agents for proactive customer re-engagement and end-to-end customer journey optimization.
Multi-agent AI workflows are limited not only by model quality but by token cost, latency, and context-window quality. This paper presents a practitioner framework for token optimization and context-window management, grounded in an internal production dashboard that extracts structured work items from meetings, email, and chat with LLMs and routes summaries across workstreams. Six patterns are described: context stratification, fetch-once/process-locally architecture, schema-contracted prompts, token-aware fallback chains, semantic caching, and inter-agent communication compression. In production they cut measured cold-load latency to 61-116 seconds (six timed runs) from an operational baseline of roughly 3.5-10.5 minutes, with an estimated 60-70% token reduction. It also reports a controlled context-composition study: 2,420 confirmatory trials across 11 model configurations, using 661 anonymized workplace items scored for relevance. Holding the prompt at a fixed ten items, replacing some high-relevance items with same-domain low-relevance items improves the model's relevance-score concordance on the target items, versus high-relevance items only; we call this relevance-contrast context. In the all-11 paired analysis, the 50:50 signal/noise condition improved relevance accuracy by +0.077 over the 100% condition (naive 95% CI [+0.056, +0.098], Cohen's d = 0.49, Holm-adjusted p < .001, n = 220). These cells are not independent; by the nine model families the effect is +0.084 (95% interval [+0.064, +0.103]), reported as a within-corpus descriptive comparison, not a population inference. A Fusion-of-N follow-up found that learned synthesis did not beat the mechanical set union of item IDs. The contribution is a measured engineering layer between model research and production agent practice: repeatable patterns and evaluation methods for faster, cheaper, more reliable workflows.
Stale recommendations are a pervasive challenge and a leading source of user complaints on large-scale content platforms. Items lose relevance through two primary mechanisms: supersession, where emerging updates render prior coverage stale, and relevance decay, where an item's informational value naturally diminishes over its lifecycle. Traditional countermeasures serve as crude proxies: age cutoffs poorly reflect actual relevance loss, while engagement heuristics rely on lagging signals, broadly exposing users to stale content before the system adapts. We present SDF (Supersession-Decay Filtering), a staleness filtering system fully deployed in Google Discover, a personalized recommendation feed with hundreds of millions of daily and billions of monthly active users. SDF targets both mechanisms with complementary filters, each powered by a learned model: a relational staleness model that detects supersession between item pairs, and a predicted traffic ratio (PTR) model that forecasts relevance decay from the item's content, trained on lifetime visit traffic. Applied via disjunction upstream of the ranking stage, SDF prunes stale candidates, measurably reducing downstream serving costs. Online experiments demonstrate that these filters significantly reduce the prevalence of stale content while improving user engagement. Over a two-year production deployment, user-filed staleness reports (in-product user feedback) declined by 54.9% relative to the pre-deployment baseline, establishing SDF as a robust and scalable paradigm for resolving content staleness at industrial scale.
Haining Xie, Xiaokai Zhou, Jiaming Yang +12cs.DB cs.AI cs.SE
Enterprise data warehouses (DWs) support business-critical analytics, but warehouse task delivery remains a complicated production process involving context retrieval, workflow configuration, code generation, platform submission, and failure diagnosis. Although large language models (LLMs) and coding agents have improved software development, they are insufficient for production DW delivery, which requires dependency-aware orchestration, lifecycle-aware artifact control, and continuous adaptation to evolving platform practices. We present SiriusDeliver, an end-to-end delivery automation agent for production warehouse task submission. SiriusDeliver integrates three components: a hierarchical delivery agent that orchestrates warehouse skills, an artifact lifecycle control module that verifies and revises artifacts before and after platform execution, and a trace-driven skill evolution mechanism that maintains reusable skills from delivery trajectories. We evaluate SiriusDeliver through offline datasets and large-scale production deployment on Tencent Cloud WeData. Offline experiments on real-world warehouse delivery cases show that SiriusDeliver improves delivery success and automation efficiency over representative baselines. During a two-month deployment across 6 business teams and 4 warehouse task types, SiriusDeliver served 3,600 monthly active users and supported 18,240 delivery sessions, achieving an 87.2% end-to-end success rate and a 73.5% autonomous submission rate. A one-month A/B test shows that SiriusDeliver reduces median delivery time from 228 to 23 minutes and engineer effort from 95 to 11 minutes, while maintaining comparable final delivery success.
Jonaid Shianifar, Blaz Mramor, Fangda Zou +5cs.AI cs.LG
Real-time bidding (RTB) ad exchanges typically forward nearly all incoming requests to demand-side platforms (DSPs), even though only a small fraction receive bids. This over-distribution weakens auction outcomes: DSPs throttle participation under compute and budget constraints, reducing the effective use of limited bidding capacity. We present a competition-aware request dispatch framework that uses distributional bid prediction and probabilistic forwarding to decide whether each request should be sent to each DSP. The system adapts per-DSP thresholds over time through lightweight policy optimization to track non-stationary market conditions. We evaluate the framework through four sequential online experiments on a production platform serving over 20 billion daily requests. A full multi-DSP deployment reduces DSP request volume under the policy by 34.2% while increasing net revenue by 4.6% (p<0.001) in a recent 14-day window after an initial DSP adaptation period. Further analysis highlights strong heterogeneity across traffic segments and reveals that aggregate metrics can be misleading. Segment-level and per-DSP analyses suggest that the policy surfaces comparative advantages among DSPs, improving monetized outcomes without increasing overall request volume.
Haonan Chen, Chu Li, Zhicheng Wang +4cs.IR cs.CL cs.CV
Multimodal representation learning is a cornerstone of modern AI. By encoding multimodal queries and targets into vectors, it powers industrial search and recommendation and underpins modern agents. Real-world platforms with complex modalities and massive-scale content, such as Douyin, Xiaohongshu, and YouTube, demand both efficiency under billion-scale indexing and fine-grained discrimination for hard matching. Existing MLLM embedding models rarely satisfy both. Contrastive models are efficient but rely on pair-level supervision too coarse for fine-grained distinctions, while CoT-based models improve discrimination through explicit generation impractical to serve online. We present Douyin Multimodal Embedding (DME), a model trained in two stages to combine both strengths. Stage 1 performs large-scale contrastive pre-training that establishes a unified multimodal embedding space with broad modality and task coverage. Stage 2 supplements semantic sufficiency, the property that an embedding is grounded in retrieval-relevant evidence and preserves fine-grained counterpart-side semantics, via two mechanisms. Evidence-Grounded Typed Latent Reasoning organizes retrieval evidence through hidden-space latent reasoning, and Cross-Conditional Reconstruction enforces counterpart-side semantics through cross-directional autoregressive reconstruction. Both act only during training and add only marginal query-side overhead, so DME serves as efficiently as a standard contrastive encoder. On MMEB-v2, DME reaches state-of-the-art results at comparable scales for its 2B and 9B variants (74.8 and 78.4), with especially strong video and visual-document tasks. In production, DME delivers a 2.92% relative gain on Douyin's in-house offline evaluation set, is deployed across Douyin scenarios such as generative, image, and AI search, and yields a 0.1% Lifetime (LT) gain in online A/B testing on Douyin search.
Production LLM agents that select from large skill libraries face a limitation that semantic relevance alone cannot resolve: a skill may match a user's topic yet be impossible to execute in the current account state. We present a deployed three-stage selection pipeline for Helpmate, Wix's customer-care assistant. First, a recall-oriented semantic matcher identifies messages related to a ten-skill domain family without consulting account state. Second, a deterministic executability gate removes candidates whose internal hard-stop conditions hold. Because the gate and the skill evaluate the same exit predicates, every blocked candidate would be unable to complete under the same account state, provided predicate parity is preserved and both checks observe fresh authoritative state. Finally, the LLM decides whether to invoke one of the remaining candidates. In a post-launch production analysis of 756.6K user messages across 267.6K conversations, semantic matching retained 174,927 messages (23.1%). Within this matched stream, the gate removed 1,039,462 of 1,749,270 skill-message pairs (59.4%), saving 228.8 million skill-description tokens -- 59.1% of the post-semantic skill-description footprint. Together, semantic matching and executability gating reduced skill-description context by 90.5% relative to exposing all ten skills to every message. To test whether this pruning affects model behavior rather than context size alone, we replayed a risk-enriched cohort of 1,000 conversations with all ten skills exposed. The model selected a production-blocked skill in 78 conversations (7.8%). This counterfactual result shows that deterministic gating prevents non-executable candidates from influencing model selection, while not claiming downstream tool execution or customer-outcome effects.
Early ranking stages in recommendation systems precompute item embeddings and cache them in-model for scoring within strict latency constraints. Because this cache exists only at serving time, outside the training loop, training and serving use different item representations, a structural discrepancy that limits quality and adds operational fragility. We show that co-designing the training and serving paths removes this representation discrepancy at its source. We introduce the memory layer, an in-model key-value embedding cache co-trained with the model: the item tower writes embeddings during training and the model reads them at serving, one source of truth for item representations by construction. Always-on embeddings cover items not yet cached, so every item receives a prediction, and the design consolidates three separate trainer-to-predictor update paths into a single self-contained pipeline. Deployed in production on Instagram Reels, the memory layer raises prediction coverage from 96% to 100%, improves embedding freshness from $O(5\text{ min})$ to $O(20\text{ s})$, and narrows the training-serving Normalized Entropy (NE) gap by up to 86%, yielding over $2\times$ recall for the freshest content and a 5-6% cold start engagement lift. Because embeddings are produced during training, the system needs no separate bulk-evaluation or publish-time recomputation, cutting training-and-publish computational cost by 30% at neutral serving computational cost.
The rise of agentic AI has catalyzed a shift toward self-iterating systems, opening new frontiers for the autonomous optimization of production recommender models. This paper presents the empirical validation of a knowledge-driven autonomous agent system, deployed directly on a production large-scale Two-Tower retrieval model. By delegating the entire research lifecycle, spanning idea generation, code implementation, offline training, and metric evaluation, to a continuous closed-loop autonomous framework, the agent system executed over 40 completed autonomous training runs from scratch. Executing these runs under rigorous production-scale evaluations, the system systematically navigated hidden architectural bottlenecks on the latest production model to achieve a breakthrough ~20% relative improvement in NDCG, a gain that translated directly to a +3.77% increase in user satisfaction in live production traffic. Furthermore, the deployment exposed critical vulnerabilities in standard evaluation protocols, as the agent system autonomously discovered reward-hacking shortcuts. These findings prove that an autonomous pipeline can dramatically accelerate the pace of machine learning research and stress-test the rigorousness of underlying experimental infrastructure, while also exposing novel challenges such as reward hacking and redundant exploration of failed hypotheses.
Yuge Zhang, Yuanxing Zhang, Yichao Jin +7cs.CR cs.AI cs.LG
Money mule accounts are critical facilitators of financial fraud, yet detecting them at scale remains challenging due to the heterogeneous nature of transactional and behavioural data. We present an end-to-end pipeline for customer-level mule detection comprising three stages: (1) a LightGBM classifier trained on 280 engineered features spanning transaction patterns, account demographics, network topology, and temporal behaviour; (2) a TreeSHAP attribution layer that decomposes each prediction into feature contributions; and (3) a large language model (LLM) module that converts SHAP attributions into analyst-facing natural-language narratives. We evaluate across three open-weight LLM families and assess explanation quality through analyst feedback. In a live production deployment, the system achieves a yield rate of 89%, up from 61% under the incumbent rule-based system, with monthly alert volume expanding from 211 to 302, reflecting broader true-positive coverage rather than increased noise. This corresponds to a 60% incremental adverse detection beyond existing review workflows, substantially outperforming the rule-based approach. Qualitative feedback from analysts indicates that LLM-generated narratives reduce cognitive load during alert triage. We further discuss implications of deploying LLM-augmented explainability in regulated financial environments.
Optimizing large-scale retrieval hinges on the ability to efficiently surface candidates across diverse content tiers. However, to capture segments such as fresh and long-tail content, modern systems typically resort to a fragmented "zoo" of specialized retrieval models. This operational complexity is attributed to a fundamental challenge in heterogeneous retrieval systems, the Scaling Bias of Heterogeneity, where model capacity gains do not apply equally across diverse content tiers. To bridge this gap, we propose MESH as a unified retrieval scaling framework that mitigates this bias through a modularized architecture integrated with gated bias correction. By partitioning the feature space into independent domains, MESH enforces a structural inductive bias that reduces interference between sparse-item signals and high-frequency engagement features. This protected gradient path leads to improved scaling behavior for sparse content, empirically validated by a 14 times improvement in the power-law scaling exponent for fresh items. In online evaluations on Pinterest's Related Pins platform, a billion scale item-to-item recommendation system, these improvements translate into a +5.5% lift in fresh-item repins, alongside with 55% improvement in funnel efficiency and +0.46% improvement in user retention. Finally, our asynchronous serving strategy ensures production viability by delivering a 2.87 times improvement in system throughput. Our findings suggest MESH as a promising paradigm for consolidating fragmented retrieval infrastructures into more scalable and ecosystem-aware backbones.
Nikita Rusakov, Vladislav Meshkov, Konstantin Zorin +4cs.LG cs.AI
Predictive modeling is a core component of modern financial services, where a wide range of tasks are traditionally addressed using separate models trained on manually engineered tabular features. This task-specific approach limits reuse and makes it difficult to fully exploit heterogeneous data sources such as transaction histories and digital interaction signals. In this paper, we present an approach based on pretraining a foundation transformer model on multimodal sequences of user events. Events from multiple data sources are unified into a single chronological sequence, enabling early fusion of heterogeneous modalities and learning of general-purpose representations via a next-event prediction objective. These representations are combined with existing engineered user features, on top of which lightweight neural models are trained for multiple downstream tasks. The proposed system outperforms traditional task-specific models while reducing development overhead. The approach was deployed in production at one of the biggest banks in Eastern Europe, resulting in measurable improvements in business metrics.
In large-scale ride-hailing, hold control is a critical mechanism for improving passenger-driver experience. By selectively delaying certain driver-order pairs, the system waits for better opportunities, reduces cancellations, and mitigates wasted driver effort. However, existing industrial hold strategies often rely on heuristic thresholding over multiple predictive models, which can be brittle under non-stationary traffic and hard to optimize for multi-objective experience signals. We propose EXHOLD, a deployable two-stage framework decoupling experience-aware pair assessment from hold-time execution. In Stage I, we learn a decision model assigning each driver-order pair to discrete, interpretable experience tiers by optimizing a unified objective that aggregates satisfaction signals across the matching funnel. In Stage II, we solve for a monotone hold-time schedule via constrained optimization over empirical quantiles. This explicitly enforces service guardrails bounding the unnecessary holding of promising matches while maximizing overall experience improvement. We evaluate EXHOLD through randomized A/B experiments in DiDi's production system in Brazil. Results show consistent gains in marketplace efficiency and experience: EXHOLD increases trip completion and driver income, significantly reduces passenger cancellations, and improves funnel efficiency. Ablations and behavioral analyses confirm both stages are essential and that the policy makes calibrated decisions under spatiotemporal heterogeneity. EXHOLD is currently deployed, serving production traffic in Brazil.
Ning Liu, P Aditya Sreekar, Kalle Kujanpää +11cs.LG cs.AI
Warehouse operations are governed by Standard Operating Procedures (SOPs) that encode complex, multi-system decision logic, which must be executed reliably under strict time constraints, yet LLM agents lack mechanisms to enforce procedural compliance and degrade under the context overload full SOP specifications introduce. We present Eluna, a production-deployed agentic system for reliable SOP execution. Eluna is a graph-guided, multi-agent framework that encodes SOPs as directed acyclic graphs with progressive disclosure and delegates independent tasks to parallel sub-agents, each with persistent code execution and live data access. To meet production latency and accuracy needs, we use asymmetric episodic distillation where a strong teacher is improved through episodic error memories, then a smaller student is fine-tuned on the corrected trajectories with memory stripped, internalizing corrections without inference-time overhead. On a 13-task benchmark and two production applications, our fine-tuned models match or exceed their teacher, beat all larger off-the-shelf baselines, and reach 94% expert agreement on the ticket processing application.
Kalle Kujanpää, Ning Liu, Shahnawaz Alam +4cs.CL cs.LG cs.SE
Production LLM agents often waste latency and reliability by regenerating code for the same procedural steps on every request. We replace this inference-time coding loop with an agentic tool-making pipeline that compiles repeated SOP steps into validated, versioned tools before deployment. The tool-maker grounds synthesis in the live environment as it collects execution traces, observes backend schemas and values, generates candidate tools, and repairs them against labeled cases. At runtime, the production agent calls these tools directly and falls back to code generation only when needed. We deploy the approach in a Fulfillment Center alarm-triage system, where an agent diagnoses alarms against a 44-node SOP over heterogeneous metric backends. In production, tool calls reduce p50 latency by 42%. On 1,500 historical alarms, they reduce end-to-end error rate by up to 53% by suppressing run-to-run variance in repeated steps. Because tools return compact structured verdicts, they also enable a simpler direct-call architecture, reducing p50 latency by a further 62% in a controlled ablation. Versioned tools also improve auditability and expose specification gaps and upstream data drift. Our results show that self-evolving agents can make industrial LLM systems faster, more reliable, and easier to operate.
Yuming Liu, Hongye Yang, Harrison Zhao +5cs.IR cs.CV cs.LG cs.MM
We present NEXT (Next-interest EXploration Transformer), a reasoning-driven video recommendation framework that reasons over the video a user has just watched, infers the viewer's next intent, and retrieves concrete follow-up videos. Explicit continuations such as episodes are linked directly; implicit cases are handled by generating intent queries and searching for matching candidates. This Item-to-Intent-to-Item formulation produces directed recommendations beyond co-engagement correlation or semantic similarity. To make this framework reliable at scale, we train NEXT-8B, a purpose-trained 8B vision-language model with a three-stage recipe: Perception-Enhanced Reinforcement Learning for query-agnostic evidence extraction, Distribution-Aligned Supervised Fine-Tuning over real and synthetic visual QA mixtures, and Group Relative Policy Optimization for last-mile alignment. NEXT-8B achieves the best single-model DocVQA performance, ranking second overall only behind a multi-agent system while surpassing a substantially larger 200B+ scale model, and improves next-intent logic-wise quality by 3.3% over the base model in a task-specific LLM-as-a-judge evaluation. We deploy NEXT as an additional retrieval path in a large-scale social media recommendation system and observe statistically significant production gains, including +0.53% watch time and +0.51% distinct video exposure. Overall, NEXT shows that a carefully trained compact vision-language model can serve as a practical reasoning engine for next-interest exploration at production scale.
Pricing automation in large-scale tourism is challenging because travel orders are highly unstructured, while pricing policies are complex, rapidly evolving, and inherently open-ended. Traditional rule engines are brittle and costly to maintain, whereas unconstrained LLM agents lack the reliability and auditability required for financial decisions. We present a production-grade LLM-powered pricing system with a strict decision boundary: LLMs perform structured extraction and bounded policy/path selection, while all numeric pricing, including total-price computation, is executed deterministically. Policies are compiled into interpretable condition trees, enabling open-ended support for new clauses and evolving rules without code changes, while exposing auditable artifacts for human-in-the-loop control. Periodic fine-tuning on logged traces further improves tree induction and path matching. Deployed at a municipal state-owned tourism enterprise across 7 scenic sites and 12 business categories with 1,500+ operators and 1,000+ active policies, the system processed 3,960 orders in six months, reduced the order management team from 15-20 to 3, and cut per-order handling time from 10 minutes to <2 minutes.
Sequence modeling has become increasingly popular in recommendation and ranking algorithms, owing to its capacity to model users' historical behaviors and infer user intentions. Despite its theoretical simplicity, the practical deployment of a sequence model in production is non-trivial due to complexity of the sequence and sparse labels. For example, in Airbnb, guest sequences are often long, exploratory and complex, and we focus on booking labels, which are sparse. As such, we are often required to make various design decisions regarding data and modeling to strike a balance between effectiveness and scalability. This work delved into these production challenges and deployed JourneyFormer, a sequence modeling solution for search ranking at Airbnb. We detail crucial design considerations, covering aspects such as guest event selection, ID embeddings, model architecture, and label attribution. Additionally, we describe several tailored strategies to accelerate model training and inference. JourneyFormer has been successfully deployed within Airbnb's production, where its effectiveness and impact have been evidenced not only by improved offline ranking metrics but also by significant gains in key business metrics through online A/B testing across 2 production surfaces.
Dispatch in three-sided marketplaces provides a natural setting for reinforcement learning from world feedback: decisions are evaluated by delayed operational outcomes such as delivery speed, courier utilization, and merchant congestion. We present a deployed reinforcement learning system at DoorDash that adapts dispatch objective weights in a large-scale food-delivery marketplace using delayed signals. Rather than replacing the combinatorial assignment optimizer, a store-level policy learned from logged marketplace data selects a discrete multiplier that shifts the dispatch optimizer's tradeoff between delivery quality and batching efficiency. This interface enables offline policy learning under noisy, delayed, and coupled feedback while preserving production feasibility constraints and operational safeguards. We train a shared value function using centralized offline data and decentralized store-level execution, with Double Q-learning targets and a conservative regularizer to reduce out-of-distribution value overestimation. In a production switchback experiment, the offline-trained policy increases batching and reduces courier-side time costs without degrading customer-facing delivery quality. Results illustrate how world feedback from a live economic and logistics system can be used to safely adapt decision policies online.
Aman Gupta, Kevin Rossell, Edesio Alcobaça +8cs.CL
The rapid rise in LLM capabilities has made AI agents increasingly viable across a broad range of tasks. Among the most promising applications is building production-ready customer-facing agents, a challenge that demands coordinated excellence in evaluation methodology, context engineering, training, and online measurement. Yet these critical pillars are typically developed in isolation, creating blind spots that only surface after deployment. In this paper, we present a unified framework that bridges offline development with online impact for customer support AI agents at Nubank, a company with 100M+ users. Our approach integrates several key components: (1) structured context engineering tailored to customer support agents, (2) systematic human-in-the-loop prompt iteration, (3) rigorous LLM judge evaluation with measured inter-rater agreement and GEPA optimization for consistency, and (4) ideation-to-production validation. A central insight is that evaluation-pipeline quality directly determines iteration velocity. We present results from five production deployments spanning distinct domains: card delivery, debt management, credit-limit support, card management, and product explanation. These deployments deliver consistent customer-satisfaction gains while substantially accelerating iteration. In our card-delivery deployment, large-scale A/B testing yields a 37 percentage-point improvement in AI transactional Net Promoter Score and a 29 percentage-point gain in self-service rate over prior agent variants, alongside a strong correlation between offline simulation metrics and online outcomes, demonstrating that eval-driven development reliably predicts production impact. On most use cases, AI satisfaction reaches within a few percentage points of expert human agents.
Large Language Model (LLM) agents are increasingly used in real-world products, where personalized and context-aware user interactions are essential. A central enabler of such capabilities is the agent's long-term semantic memory system, which extracts implicit and explicit signals from noisy longitudinal behavioral data, stores them in a structured form, and supports low-latency retrieval. Building industrial-grade long-term memory for LLM agents raises five challenges: scalability, low-latency retrieval, privacy constraints, cross-domain generalizability, and observability. We introduce the Hierarchical Long-Term Semantic Memory (HLTM) framework, which organizes textual data into a schema-aligned memory tree that captures semantic knowledge at multiple levels of granularity, enabling scalable ingestion, privacy-aware storage, low-latency retrieval, and transparent provenance; HLTM further incorporates an adaptation mechanism to generalize across diverse use cases. Extensive evaluations on LinkedIn's Hiring Assistant show that HLTM improves answer correctness and retrieval F1 significantly by more than 10%, while significantly advancing the Pareto frontier between query and indexing latency. HLTM has been deployed in LinkedIn's Hiring Assistant to power core personalization features in production hiring workflows.
Modern enterprise AI applications increasingly rely on compound AI systems - architectures that compose multiple models, retrievers, and tools to accomplish complex tasks. Deploying such systems in production demands inference infrastructure that can efficiently serve concurrent, heterogeneous model invocations while maintaining cost-effectiveness and low latency. This paper presents a production deployment study of a modular, platform-agnostic inference architecture developed at Salesforce to support compound AI use cases including Agentforce (autonomous AI agents) and ApexGuru (AI-powered code analysis). The system integrates serverless execution, dynamic autoscaling, and MLOps pipelines to deliver consistent low-latency inference across multi-component agent workflows. We report production results demonstrating over 50% reduction in tail latency (P95), up to 3.9x throughput improvement, and 30 to 40% cost savings compared to prior static deployments. We further present a novel analysis of compound-system-specific challenges including multi-model fan-out overhead, cascading cold-start propagation, and heterogeneous scaling dynamics that emerge uniquely when serving agentic workloads. Through detailed case studies and operational lessons, we illustrate how the architecture enables compound AI systems to scale model invocations in parallel, handle bursty multi-agent workloads, and support rapid model iteration - capabilities essential for operationalizing agentic AI at enterprise scale.