Union.ai

📊 We benchmarked multimodal document retrieval across three approaches. The results were humbling 😅, thanks to Muhammad Adil Fayyaz 's work on creating a high performance benchmarking framework. SigLIP — a SOTA vision-language model — scored Recall@5 of 0.022 on financial documents. BM25 scored 0.142. Keyword search beat a neural model by 6x. No fine-tuning. No embeddings. Just text. The takeaway isn't that neural models are bad. It's that running one experiment cheaply is worth more than reasoning confidently about which approach should win. We built a reusable benchmarking pipeline on Flyte 2 to make that easy — task-level caching, warm GPU containers, dynamic batching. Swap in a new model, but don't recompute the index. Full breakdown of the architecture and results 👇 https://lnkd.in/ePUdVM95 #MachineLearning #RAG #Flyte #DocumentRetrieval

🔥 Stop reaching for a framework you'll regret. 🔀 Most Python devs building AI workflows eventually hit the same wall: things get complex, and suddenly the team is evaluating orchestration frameworks, learning DSLs, and debugging abstractions nobody fully understands. But what if the standard library was the DSL you needed all along? 🐍I'm speaking at PyCon 2026 on exactly this: how asyncio + container orchestration (think Kubernetes) gives you everything you need to build production-grade AI workflows — without giving up readability, debuggability, or your Python brain. We'll cover: → Structured concurrency patterns for coordinating LLM calls → Parallel tool execution with real backpressure and timeout handling → Streaming responses and rate-limited API flows — no black-box DSL required → Failure isolation and retries that don't leak complexity upward The pitch is simple: plain Python scales further than you think. And when you do hit a ceiling, you can keep the expressiveness of Python and scale with compute orchestration that works seamlessly with it. If you're shipping AI agents or workflows to production and wondering whether your framework is pulling its weight, come find out what it looks like to let it go. 📍 PyCon 2026 Friday May 15th, 2026 11am -11.30am PST See you there! #pycon2026 #python #asyncio #aiengineering #mlops #kubernetes