AI engineering after the demo.
MLOps.community sits at the practical edge of AI engineering, where agents, RAG, GPUs, and model serving meet real operational constraints. The tone is conversational but technically dense, built around people comparing what works, what breaks, and what still feels unresolved.
PRESS PLAY
MLOps.community is a technical conversation show about the hard part of AI: making it work after the demo. The emphasis stays close to production. Recent conversations cover coding agents, LLM latency, distributed training, vector search, agentic marketplaces, Databricks roundtables, and the changing career path of software engineers. The show assumes technical fluency. Acronyms arrive without much ceremony, and guests are allowed to think out loud through architecture, workflow, and tradeoffs. That looseness is part of the appeal. A conversation about TensorRT-LLM becomes a practical look at GPU efficiency, cold starts, and leaders who have to learn AI beside their teams. A discussion of agent teams turns screenshots, pixel diffs, and QA reports into a management problem: scarce human attention, better reporting, and agents that know when to ask for help. Episodes about software engineering keep returning to the same uncomfortable question. If tools make junior work easier to automate, the industry still has to train people into senior judgment. The show does not treat that as a slogan. It lets the concern sit beside jokes, side quests, and practical examples from code review, pair programming, and planning tools. Infrastructure episodes go deep. GPU starvation, local caching, vector spaces, graph wiring, metadata layers, search behavior, and production reliability all get room. Product conversations widen the lens. In the agentic marketplace episode, classifieds become a way to think about personal agents, structured choices, dashboards, chat interfaces, and the cost of wrong assumptions. The tone is curious and sometimes excited, but rarely polished into a clean keynote shape. Guests repeat themselves, qualify claims, and circle back. That makes the show feel closer to a technical hallway conversation than a finished essay. MLOps.community is strongest when it faces the gap between speed and responsibility: agents can move fast, but engineering teams still need trust, evaluation, cost control, and ways to stop bad work before it reaches users.
Made for: For software engineers, ML practitioners, engineering leads, and AI builders who already speak the production vocabulary. It fits listeners looking for grounded conversations about agents, infrastructure, reliability, and workflow change.
What sets it apart: The show lets technical conversations stay messy enough to be useful, with long explanations instead of tidy hype cycles. Its center of gravity is production reality: latency, cost, QA, GPUs, trust, and human oversight.
Relaxed Conversations around getting AI into production, whatever shape that may come in (agentic, traditional ML, LLMs, Vibes, etc)
Based on 6 episodes we listened to · September 2026
MLOps.community makes production AI legible by connecting agent behavior to infrastructure, reliability, performance and the daily mechanics of building software.
MLOps.community follows AI from prototype enthusiasm into the awkward realities of production. The host moves through technical conversations and a roundtable on LLM serving, durable execution, GPU performance, coding-agent workflows, marketplace design and vector search.
From overheating GPUs to agents resolving merge conflicts, your AI prototype meets its consequences.
Press play if you want
Short reviews from the PlayNext desk, based on the episodes we processed.
A focused tour of the feature space behind geometric vector search.
Arrowspace pushes vector search beyond the usual idea of item-to-item similarity. The guest follows the project from early experiments into a published library, with an eye toward making vector search both more accessible and more capable for text embeddings and highly connected…
A practical argument for guided agents in marketplaces.
This discussion pulls agentic marketplaces back toward product reality by grounding the idea in OLX's motors, real estate, and classifieds businesses.
A clear look at durable execution when agents, timers, and failure recovery meet.
The episode makes durable execution feel concrete, grounding the idea in failures engineers actually meet: flaky servers, overloaded services, down APIs, rate limits, and long-running work that has to keep going.
A grounded look at what it really takes to make GPUs run fast without fooling yourself.
The episode treats performance tuning as a stack problem, not a slogan, moving from kernel-level behavior to application pressure and back down to the hardware underneath.
A practical look at when LLM serving needs its own architecture.
Focus stays on serving LLMs in production, and it draws the line where a prototype with real users starts to run into cost, GPU pressure, and fit.
A grounded Databricks review where managed convenience meets the stubborn realities of MLOps.
Databricks comes through as a practical case study in the tradeoffs behind managed MLOps. Maria, an MLOps tech lead at Aho Del Geise and co-founder of Marvelous MLOps, maps where the platform smooths the path and where its built-in assumptions can start to pinch.
A practical look at agents, diffs, parallel work trees, and provider flexibility.
The episode stays practical, centering on ways to narrow an agent's attention through pixel diffs, zooming in on the actual error, and reports that surface the most important points.
Power Platform news with practical static.
The episode explores strategies for managing teams of AI agents in software development, focusing on reducing human intervention.
Maria Vechtomova, an MLOps tech lead and co-founder of Marvelous MLOps, discusses her deep expertise with Databricks in this episode.
James, global head of AI enablement at MetLife, explains how he uses FinOps Foundation thinking to balance model cost, speed and accuracy, and the host proposes a 'cave bench' to test how far caveman-style terse prompting saves tokens.
James says MCP inefficiency often comes from tool design: one customer lookup repeated 100 times can become a single aggregate call.
An MCP server can deliver agent skills when a task needs them, rather than requiring a separate skills repository.
This episode discusses how AI tools like Cursor are creating a training gap for junior software engineers, who fear being replaced by automation while struggling to gain the mentorship and experience needed to advance to senior roles.
This episode discusses how using Tensor RT LLM reduced latency by up to 70% in production, while exploring the speaker's journey into AI leadership and the importance of adapting to market shifts in technology.
This episode explores how to effectively manage teams of AI agents by teaching them to report efficiently and automate tasks like visual regression testing to reduce human oversight.
This episode discusses technical solutions for GPU starvation in distributed machine learning training and explores the evolving role and cost considerations of software engineers in the AI era.
The hosts discuss spec driven development and share personal experiences from the recent coding agent conference, including unexpected travel and community connections.
A Databricks roundtable discussion on transitioning AI agents from experimentation to production, addressing challenges, safety concerns, and the excitement around rapid advancements in software engineering.
MLOps.community’s Arrowspace episode centers on graph wiring, vector search, and epiplexity as retrieval tools for ML operations and LLM operations.
Pedro and Done spend the episode working through how to turn agent visions into real product execution.
The host and a guest talk through durable execution as a way to keep software moving when failures show up.
The episode follows a technical conversation about performance optimization across CUDA and GPUs.
The host opens by saying MLOps.community is kicking off its first panel of 2026.
This episode discusses the massive infrastructure investment in AI and machine learning datacenters, exploring how systems like Netbox help manage and track this hypergrowth.
This episode features a fireside chat with Ricky Doar, VP of Solutions at Cursor, discussing the AI tool's rapid growth, enterprise adoption, and emerging patterns in how developers are using it effectively.
The CEO of Chroma discusses context engineering, context rot, and agentic search, highlighting challenges with large context windows in AI models and misinformation in the AI space.
The VAST Data CTO discusses the challenges of scaling AI infrastructure from personal projects to enterprise-level systems, highlighting rising expectations and messy data realities.