Uber

Uber details the Software Factory it runs internally: four layers of agent usage from specialized to general, a cost equation for decomposing agent spend, how each term is measured, and the routing and optimization work that flattened total AI spend while usage multiplied.

Uber built Genie, a RAG-based Slack copilot that answers on-call engineering questions from fragmented internal docs, choosing RAG over fine-tuning to cut time-to-launch and detailing their approach to hallucination and data-source protection.

Uber's product organization reports on a year of AI prototyping across teams: interactive flows now precede the PRD, roughly 40% of hackathon submissions used a prototyping tool, and the patterns that kept recurring when prototypes replaced written debate.

Uber built Finch, a Slack AI agent that uses RAG and self-querying to turn plain-language finance questions into secure data retrieval, replacing manual SQL lookups.

Uber's FixrLeak combines AST analysis with GenAI to auto-detect and rewrite Java resource leaks into try-with-resources fixes, cutting manual remediation across their large Java codebase.

Uber's engineering account of Michelangelo, an internal ML-as-a-service platform built to replace one-off model pipelines. Covers the full workflow from data management and training through deployment, online serving, and prediction monitoring, and the reproducibility problems that forced it.

Uber Eats' Cart Assistant turns a text prompt or a photo of a recipe into a draft grocery cart through a multi-prompt state graph. LLMs handle intent planning and relevance judging; deterministic APIs own inventory retrieval, pricing and cart construction, with guardrails between them.

Uber AI Solutions explains Requirement Adherence, a system that uses LLMs to extract atomic rules from client SOP documents and enforce them live inside its labeling tool, catching text-labeling errors before submission.

Uber AI Labs details Go-Explore, an algorithm that archives promising states and returns to them before exploring further, cracking Montezuma's Revenge and Pitfall where standard deep RL agents had stalled near zero.

Uber built a GenAI-powered invoice processing system using ML and NLP to automate document extraction, replacing much of its manual and RPA-based workflow.

Uber rebuilt traffic forecasting as DeepETT, a graph-aware transformer serving over 2M real-time predictions per second across a 100M-segment road graph. Walks through the modeling and serving decisions behind a 6% long-trip ETA accuracy gain and 19% more forecast variance explained.

Uber describes its GenAI Gateway, a unified internal platform mirroring OpenAI's HTTP/JSON API that lets 60+ LLM use cases access OpenAI, Vertex AI, and Uber-hosted models through one interface, with centralized security review and budget management.

Uber engineers recount building a legal redlining agent that works inside Microsoft Word. They cover problem framing, earning lawyers' trust, improving it from feedback, and how they would rebuild it with current agent harnesses.

Uber explains why it moved from Lucene's HNSW to OpenSearch for billion-scale vector search, citing algorithm flexibility and a FAISS-based GPU acceleration path.

Part one of Uber's move of machine learning workloads onto Kubernetes with Ray. It covers what forced the migration, including engineers hand selecting clusters and GPU SKUs and static configs that left some clusters saturated and others idle, plus the job management layer built to replace that.

Uber details its MCP Gateway: a proxy layer translating MCP to existing protocols, a registry and API-crawling discovery layer, and a control plane for authoring, hosting over 800 MCP servers and 5,000 tools.

Uber Engineering describes an agentic system that autonomously removes stale feature-flag code, handling boolean logic, helpers, call sites, configs and tests, with validation to keep changes safe across large repositories.

Uber AI Labs' Paired Open-Ended Trailblazer grows a population of environments and agents together, periodically transferring an agent from one environment into another. Tasks that resist direct optimization become solvable through automatically discovered stepping stones, and the best curriculum is

Uber details the stack behind its in-house LLM training: on-prem A100 and cloud H100 fleets, DeepSpeed and FSDP for scale, PEFT and LoRA for fine-tuning, Ray and vLLM for throughput, and where domain fine-tunes beat calling a frontier API for Eats search and tagging.

Uber put ML validation inside its own labeling tool to catch video bounding box errors automatically: ID swaps, position jumps, freezes, and scale distortions, scored by confidence and grouped so a reviewer fixes clusters instead of re-checking every frame.

Uber AI Labs open-sources Pyro, a PyTorch-based deep probabilistic programming language merging Bayesian modeling with neural networks, detailing why they built it and who benefits: probabilistic modelers wanting to scale to large datasets and deep architectures.

Uber's engineering team details how they scaled AI/ML infrastructure for generative AI: hybrid on-prem/cloud GPU allocation, Model Flops Utilization as the core efficiency metric, and 99% uptime SLAs for long-running training jobs.

Uber's ML team walks through rebuilding the Uber Eats homefeed ranker: a DCNv2 path for sparse and dense signals running alongside a transformer encoder over chronological click and order sequences, with target-aware attention, near-real-time features, and a roadmap toward generative recommendation.

Uber describes the retrieval and evaluation changes that improved its internal RAG copilot Genie's answer accuracy by a relative 27% on security/privacy queries.