QCon

Complex transitions call for architecture and cross-team alignment. Javier Fernandez-Ivern, Staff Software Engineer @Netflix, hosts QCon San Francisco's Staff+ Engineering Skills track on alignment and technical direction. Explore the track: buff.ly/IPzPGZ7

Audio LLMs combine recognition, generation, translation, and dialogue. Tara Sainath examines these capabilities in her QCon AI New York 2026 keynote. Read more here: buff.ly/XpG4iH2

A patch may take time; runtime mitigation can help in the meantime. Liz Rice, Chief Open Source Officer @Isovalent at Cisco, will demonstrate eBPF and Tetragon at QCon London 2027. Read the abstract: buff.ly/xM13s0z

Blackbaud’s lakehouse prompted a shift toward clearer data ownership. Shannon McKenzie, Sr Principal Engineer @Blackbaud, will examine the move to owned data products at QCon San Francisco 2026. Read the abstract: buff.ly/G91kXU6

Coding tasks can run in parallel, with candidates checked against tests. Meryem Arik, Co-Founder & CEO @Doubleword explores swarms of open models at QCon AI New York. Read the abstract: buff.ly/XlKWKuB

When a codebase is predictable, agents can generate more correct code. Jo Kelly-Fenton, Engineer at Spotify, explains the compounding effect: more correct code is cheaper to verify automatically, which reduces review load and creates more capacity to keep improving the codebase.

Building an evaluation dataset can look like a large upfront investment, which is one reason teams defer it.

Lan Chu, AI Tech Lead and Senior Data Scientist, recommends starting with a small set of real user queries covering the key production cases. They can give more useful signal than large batches of synthetic examples.

A question about dependency injection frameworks on the company Slack came back with 342 replies.

Jo Kelly-Fenton, Engineer at Spotify, explains how a five-person advisory board was given responsibility for the call, turning a company-wide debate into a smaller, explicit decision-making process.

Running an AI system a few times and getting reasonable-looking output isn't the same as evaluating it.

Lan Chu, AI Tech Lead and Senior Data Scientist, describes two phases: a controlled dataset to test quality before launch, then production signals to show where the evaluation set and the system need to evolve.

Airplanes can largely fly themselves, yet pilots still monitor the system and judge when something isn't safe.

Aleksandar Mitic, Software Engineer at Spotify, uses that analogy for reviewing agent-generated PRs: the reviewer monitors the evidence and decides whether the change is safe to merge, rather than recreating every step.

A production RAG query about Mexico's 2026 monetary policy returned perfectly relevant results. Every one was from 2020 or 2021.

Lan Chu, AI Tech Lead and Senior Data Scientist, explains the missing signal: recency wasn't part of the scoring, so the retrieval system never represented the query's time intent.

Latency, consistency, partial failure, scaling, and observability shape whether a distributed system can be trusted. Review the Distributed Systems in Production track at QCon San Francisco 2026: buff.ly/vdUEEn4

Aggregations such as sum and count don't depend on the order in which data is processed.

Nikhil Simha, Co-Founder & CTO at zipline.ai, explains how Chronon uses that property for rolling-window aggregations: split data across machines, aggregate independently, merge the results. The outcome doesn't change.

Don't try to recreate CI inside a coding agent runtime. Use the CI system you already have.

Jo Kelly-Fenton, Engineer at Spotify, shares the lesson: the more dimensions of correctness you can verify automatically, the more evidence the agent has that its change actually works.

A counterintuitive lesson from a RAG build: when documents are small enough and cost isn't a constraint, skipping chunking can remove significant engineering complexity.

Lan Chu, AI Tech Lead and Senior Data Scientist, explains when not to chunk, and why structured chunking is a strong first approach when you do need it.

When Apache Fluss elects a new leader, recovery has to account for both confirmed and in-flight data.

Anton Borisov, Principal Data Architect at Fresha, breaks down the sequence: truncate to the last confirmed high water mark, restore from a RocksDB snapshot, then replay the in-flight data into the prewrite buffer.

Building a new ML feature traditionally means stitching together Spark, BigQuery, Flink, Redis, Kubernetes, and Airflow, repeated for every new signal.

Nikhil Simha, Co-Founder & CTO at zipline.ai, explains how that growing burden contributed to data science and ML engineering becoming separate, large teams. Chronon was built to move that repeated infrastructure behind a common system.

Once AI is part of a critical system, model quality is only one control. Evals, inference, reliability, and autonomous workflows matter too. The Engineering AI Systems track at QCon San Francisco connects those decisions through six production case studies.

Giving every engineer a coding agent does not create shared team capability. Read Brian Yang’s QCon AI New York session abstract: buff.ly/irPyeXP