--- title: "Inside Uber’s GenAI Stack" newsletter: "MLOps Community" date: 2025-07-10 source: https://aaif.live/newsletters/mlopscommunity/2025-07-10-inside-uber-s-genai-stack --- # Inside Uber’s GenAI Stack *Plus, building AI-native systems based on what LLMs actually do, mini sunmit on building AI that doesn't break, building an MCP server for hacker news, Hidden Gems, and ML Confessions.* *MLOps Community — Agentic AI Foundation, 2025-07-10* I asked on LinkedIn how folks are evaluating agentic systems - especially how it compares to more standard eval workflows. Got some great replies, but curious here too; if you're working on this (or know someone who is), let me know. LinkedIn thread here [https://go.mlops.community/IntroJul10]. What’s behind every good eval? Good data. Show you’re part of the setup - the AI Training Data [https://go.mlops.community/TrainingTJul10] tee is 30% off this week. ## Inside Uber’s AI Revolution - Everything about how they use AI/ML Some teams are building GenAI to create images - Kai’s team built an artist. Michelangelo now powers thousands of ML models across Uber, from fraud detection to pricing and rider-driver matching. Most teams rely on standard templates, but advanced users drop down to Uniflow and Ray for full control. As GenAI workloads grow, the platform’s infra has evolved to support much larger models. That’s meant adding: * DeepSpeed for multi-GPU training * Ray for orchestrating complex workflows * Triton for scalable model serving Open-source plans are in motion, starting with Uniflow. Click below to listen - this one’s a work of art. Video [https://go.mlops.community/pkwjul10] || Spotify [https://go.mlops.community/skwjul10] || Apple [https://go.mlops.community/akwjul10] [https://go.mlops.community/akwjul10](https://go.mlops.community/akwjul10) [https://go.mlops.community/pkwjul10](https://go.mlops.community/pkwjul10) [https://go.mlops.community/skwjul10](https://go.mlops.community/skwjul10) ## Hidden Gems ## A New Way of Building with AI A talk Jiquan might give at next week’s conference: how he finally cracked Google Docs’ infuriating table API. Beyond that, he shared how Lutra builds AI-native systems by designing around what LLMs actually do well, instead of forcing them to use developer-shaped APIs. They focus on getting real work done: updating spreadsheets, syncing CRMs, parsing invoices - all with code-first orchestration that scales. Their sandboxed runtime helps make that possible by: * Persisting state across sessions * Estimating cost and runtime before launch * Understanding schemas to avoid integration failures Just click to watch - no APIs to wrangle. Video [https://go.mlops.community/pjn10jul] || Spotify [https://go.mlops.community/sjn10jul] || Apple [https://go.mlops.community/ajn10jul] [https://go.mlops.community/ajn10jul](https://go.mlops.community/ajn10jul) [https://go.mlops.community/pjn10jul](https://go.mlops.community/pjn10jul) [https://go.mlops.community/sjn10jul](https://go.mlops.community/sjn10jul) ## Job of the Week [https://go.mlops.community/ROWJob10Jul](https://go.mlops.community/ROWJob10Jul) [https://go.mlops.community/MLConfess](https://go.mlops.community/MLConfess) ## Building AI that Doesn’t Break ## MINI SUMMIT With perfect timing ahead of next week’s Agents in Production Part 2, this mini summit offered a sharp look at the real challenges behind running AI systems reliably. The talks tackled one core issue: how to make AI workflows resilient, maintainable, and safe to run in production. Each speaker moved away from brittle LLM improvisation, toward structured, dependable execution. Workflow durability came up repeatedly, with different approaches: * DBOS checkpoints function inputs and outputs to Postgres, enabling retries, sleeps, and human-in-the-loop flows without duplication. * Hera wraps Argo Workflows in a Python SDK, letting engineers define type-safe DAGs that compile to YAML. * Rasa’s “process calling” approach keeps business logic stateful and predictable across multi-turn conversations. Click below for the perfect warmup. Watch it here [https://go.mlops.community/8zjf8l] ## Build an MCP Server for Hacker News This newsletter's only twice a week, so it can be hard to stay on top of everything in between. This blog walks through building an MCP server for Hacker News using Python - a great way to let a model fetch or search live content like stories, users, or comments. Tool setup is a key focus, showing how: * Each tool is described using JSON Schema, enabling validation and autocomplete * Logic is tied to functions that fetch and format data for the model * Everything runs over stdio for integration with Claude Click below to read it and maybe you’ll build one for these newsletters next. Read it here [https://go.mlops.community/BlogJul10] [Read it here](https://go.mlops.community/BlogJul10) If your metadata tool feels like Notion for tables, it’s probably not helping anyone ship faster. Is it time to treat metadata as ops, not docs? https://go.mlops.community/DataT Working on something tricky or planning ahead? Here’s how we can help - just hit reply: * Custom workshops tailored to your company’s needs * Hiring? I know some quality folks looking for a new adventure * Want to connect with someone tackling similar problems? I can introduce you Thanks for reading, catch you next time! --- Source: https://aaif.live/newsletters/mlopscommunity/2025-07-10-inside-uber-s-genai-stack