Today's trending projects on GitHub share a common thread: infrastructure problems that have quietly blocked AI adoption at scale. Whether it's stitching sensor data into coherent 3D maps, reconciling conflicting KPI definitions across analytics tools, or bridging the gap between research papers and production deployments, the most-starred projects this weekend are attacking the unglamorous plumbing work that makes serious AI applications actually possible.
LingBot-Map: Real-Time 3D Scene Reconstruction from Video Streams
Robbyant's lingbot-map is a feed-forward 3D foundation model that reconstructs scenes from streaming video at around 20 FPS — even on sequences longer than 10,000 frames. The architecture uses a Geometric Context Transformer that combines coordinate grounding, dense geometry estimation, and long-range drift correction into a single streaming pass, avoiding the expensive iterative optimization that makes most 3D reconstruction pipelines impractical for real-time use. It ships with a browser-based Viser viewer, ONNX-backed sky masking for outdoor footage, and support for benchmarks including KITTI, Oxford Spires, and Tanks and Temples.
github.com/Robbyant/lingbot-map
Apache Ossie: A Vendor-Neutral Standard for Semantic Metadata Exchange
Apache's ossie project addresses one of the quieter frustrations in modern data stacks: the same business concept — "monthly active users," "net revenue," "conversion rate" — ends up defined differently in every tool that touches it. Ossie provides a JSON/YAML specification that acts as a single authoritative source for semantic metadata, with reference converters already built for dbt, GoodData, Polaris, and Salesforce. The practical upshot is that AI agents operating on business data can rely on consistent, validated definitions rather than silently inheriting whichever tool's interpretation happened to run last.
PostHog Adds AI Observability to Its Open-Source Analytics Platform
PostHog — the open-source product analytics platform with over 36,000 stars — has expanded its feature set to include LLM observability, letting developers capture traces, generation costs, and latency metrics for AI-powered features alongside their standard product analytics. The appeal is consolidation: instead of running a separate tool to monitor AI model behavior, PostHog integrates that signal directly into the same dashboard tracking session replays, feature flag rollouts, and error rates. It now also supports MCP for IDE integration, positioning it as a development-time observability layer as much as a product analytics tool.
AI Engineering from Scratch: A 503-Lesson Open-Source Curriculum
Rohit Ghumare's ai-engineering-from-scratch is a structured, free curriculum that walks from linear algebra through multi-agent coordination across 20 phases and roughly 320 study hours. Each lesson follows the same pattern — problem statement, mathematical derivation, hand-coded implementation, test suite, and a reusable artifact — so students end up with a portfolio of prompts, skills, agents, and MCP servers by the end rather than just theoretical knowledge. The curriculum covers Python, TypeScript, Rust, and Julia, and includes placement quizzes designed specifically for use inside Claude, Cursor, and similar AI coding assistants.