import_bengaluru

Every AI framework, dataframe library and scientific tool in Python eventually bottoms out in NumPy, yet integer division is one of the most expensive things a CPU can do. Follow a multi-year contribution journey from replacing division by a constant with multiplication and shifts, through portable SIMD on Intel, AMD and ARM, to edge cases such as INT_MIN // -1. The session also covers a regression shipped to millions of users and what benchmarks, review and fixing it revealed about real open-source performance work.

Learn how Point Blank became what it is today, then see how Triton makes it possible to write fast, low-level GPU kernels in pure Python instead of dropping into C++ and CUDA.

LLM knowledge alone is not enough for many real-world tasks. Using practical Python and SerpApi examples, this session shows how tools, MCP and reusable skills let agents search the web, retrieve current information and solve tasks beyond what an LLM can do on its own.

Building the abstraction was the easy part. We designed an Offline Job Orchestration Platform—a thin layer over Airflow where a DAG declares intent (data, application and runtime) instead of hand-wiring paths, SparkConf and Kubernetes manifests—and it worked almost immediately. Then it sat there. More than 600 existing DAGs stayed where they were because migrating meant each team rewriting its own code on our schedule instead of theirs. Adoption is a cost problem, not a persuasion problem, so we stopped writing migration guides and shipped the migration itself. A transpiler built on Python's standard-library ast module rewrites legacy DAGs onto the new abstraction, while a triage classifier identifies which DAGs are automatable and which functions the tool should learn next. You'll learn to write your own codemod, what ast gives you for free, where it breaks and why the differ mattered more than the generator. Takeaways: every user step is an adoption tax; trust is the bottleneck, so ship the diff that proves nothing changed; and rejected DAGs are a design review of your abstraction. A platform nobody has migrated to is only a proposal.

Start with the core ideas behind PyTorch, then explore ExecuTorch and how PyTorch models move from development to efficient inference on mobile and edge devices.

Build a retrieval-augmented generation application in Python and learn how MongoDB Vector Search stores embeddings, retrieves relevant context and grounds an LLM's answers in your own data.

A demo-led build of a minimal AI agent harness from the ground up. Unpack the agent loop, tool calling, context management and memory, then combine them into a working harness that clarifies how tools such as Claude Code and Codex work.