Partnership opportunity with Translational Imaging Innovations

It’s impressive to see how accessible computer vision has become thanks to open-source projects like YOLOv11. This video demonstrates a full PPE detection system all built using open-source tools and frameworks. What stands out is how fast and lightweight these new YOLO versions are, making real-time edge deployment possible even on small devices. https://lnkd.in/eSwKXAWF #OpenSource #YOLOv11 #ComputerVision #AI #EdgeAI #RaspberryPi #Innovation #PPE

Today we delivered an optimization for a hand tray. Beforehand, we decided to run a simulation to see what would happen if the tray were split into two. Using a combinatorial search, we built a cost function around unused instruments and let the data determine the most efficient 2-tray configuration. We didn’t tell the algorithm anything about procedure types. The crazy thing is that it produced a near-perfect split between basic soft-tissue cases and more complex bone and joint procedures. Our initial recommendation would have cut the tray size by 28%. After seeing the data, the CNO and Chief of Ortho backed the “Hand Minor” and “Hand Major” setup. It will reduce the number of instruments sent to the OR by 42% on a tray they just right-sized last year. Somehow I still get surprised by the potential impact of mente data.

New release Ultralytics v8.3.214 | Confidence in classification validation plots 📊 Adds confidence overlays to classification validation plots and hardens NaN recovery for more robust training, with cleaner logs and docs updates for SAM 3 and YOLO26. Minor updates: ✅ More reliable NaN/Inf auto-recovery across training resumes ✅ Fresh results.csv per run when not resuming for tidy experiment logs ✅ Documentation updates: SAM 3 page and YOLO26 marked Coming Soon Ultralytics v8.3.214 release notes ➡️ Read the release notes https://lnkd.in/dWscN2yc

The last post sparked questions about how GLoRA-Hive actually works. This clip shows it in action in mroe detail. Result first: in this demo we climb past 60 good units, near the ~62-unit upper bound for this scenario. The video comes from our simulated factory environment with simulated camera views of two independent machines. Each step adds a new point to the chart below, showing how output evolves over time as the agents adjust their parameters. The twist: the agents do not communicate. Through our GLoRA-Hive approach (genetic, online LoRA-style adaptation across multiple LLMs), they still develop emergent coordination, pushing toward near-optimal production. The simulator is stochastic, so the same settings do not always yield the same result. That variability makes the learning harder and closer to real operations, where noise and drift are the norm. This is one slice of what we are building. More soon.

Recreating LoRA Without Regret: Efficient Fine-Tuning for LLMs Found this interesting blog on Hugging Face TRL about LoRA (Low-Rank Adaptation) fine-tuning. Key insights: • LoRA matches train rewards of full fine-tuning with much less memory use • Approach allows fine-tuning larger models on standard hardware • Training remains stable and performance is robust across tasks • Chart shows SmolLM3 achieves competitive rewards with LoRA • Practical implications for scaling LLM training efficiently Blog: https://lnkd.in/dQ-mDkVZ #LoRA #Finetuning #LLM #GenerativeAI #TRL #HuggingFace

The ERO-DL system is set to transform cold glue application, using artificial intelligence to simplify setup tasks for operators while intelligently inspecting adhesive patterns on any surface. https://lnkd.in/eG6JJWA6

 "Is Parlant like DSPy? What's the difference?" We've been seeing this question lately. I get why people are confused, as both frameworks claim to make LLMs more consistent. But they actually solve completely different problems.  - DSPy solves: "My model outputs are incorrect."  - Parlant solves: "I can't trust my agent with our customers." But they're actually complementary. You could use DSPy to optimize specific components (like RAG retrieval) while using Parlant to enforce effective and business-aligned customer interactions.  I wrote a detailed breakdown of the differences, including:  - How each framework actually works (with code examples)  - When to use which  - How to use them together  - The difference between optimization and control  If you're building LLM applications and wondering which approach fits your needs, this should help clarify it! https://lnkd.in/d_YQ4UyE

New release Ultralytics v8.3.219 | FastSAM prompt segmentation accuracy and stability boost 🚀 FastSAM gets precise mask upsampling and built-in CLIP for text prompts, boosting segmentation quality, stability, and ease of setup. Learn more in the FastSAM docs https://lnkd.in/dFT5BkNr. Minor updates: ✅ Reliable mask upsampling for cleaner masks during prompts ✅ Integrated CLIP in `ultralytics.nn.text_model.CLIP` for consistent text–image similarity ✅ Better compatibility across PyTorch versions and image sizes Ultralytics v8.3.219 release notes ➡️ https://lnkd.in/dV3HCxrS

This Princeton University lab built out an agent evaluation harness using W&B Weave: https://lnkd.in/gd54sF6D "We utilize Weave (Weights & Biases, 2024), an open-source monitoring tool compatible with prominent LLM libraries (OpenAI, Anthropic, LiteLLM (YC W23), etc.), that automatically captures model API calls, tool calls, API parameters, and usage information. Weave unifies telemetry across providers and frameworks, giving us consistent call graphs, usage summaries, and per-call metadata without modifying agent logic. Weave requires minimal changes for agent developers to use: the HAL harness automatically initializes the Weave client at the beginning of each run and automatically tracks usage. The Weave client is initialized with a unique run identifier, and we set a weave task id attribute on downstream calls, which enables grouping LLM and tool calls by benchmark task."

What if your line caught the defects that matter, on the first pass? We created an algorithm that learns the real defect patterns from your line and generates synthetic variants. It improves any vision model by adding better data. In this carousel, we walk you through the impact from a recent project. Swipe to see it for yourself!