AI Open Source Radar — July 19, 2026

Selection: only AI-related GitHub repos with steep recent (yesterday/today) star growth. Large, plateaued, or rank-holding repos are excluded even when cumulative stars are high. Primary metric: Trendshift daily (snapshot verified at 2026-07-18 00:06 UTC via Snowflake-decoding of the page’s live-mention post IDs), cross-checked against each repo’s detail page for cumulative stars, language, and license. Excluded as decelerating: xai-org/grok-build (daily +9,500 → +4.3k, −55%, 17k total) and Fei-Away/Codex-Dream-Skin (+5,200 → +3.4k) — both featured yesterday and clearly cooling.

Robbyant/lingbot-map — +1.3k ⭐ today (7.8k total · #3 on the daily board)

What it does: A feed-forward 3D foundation model that reconstructs a scene’s 3D structure and the camera trajectory in real time from ordinary streaming RGB video — no depth-sensor hardware required.

How it works: At its core is a Geometric Context Transformer that unifies three things in a single streaming framework: coordinate grounding (anchor context), dense geometric cues (a pose-reference window), and long-range drift correction (trajectory memory). Built as a feed-forward network with paged KV-cache attention, it sustains about 20 FPS on 518×378 inputs and stays stable across sequences longer than 10,000 frames.

Use cases & edge: An online spatial-understanding layer for robotics and embodied AI — the real-time map/pose that underpins navigation, obstacle avoidance, and manipulation. It delivers SLAM-grade results from RGB alone, without costly LiDAR or depth rigs.
Why it’s rising now: After first trending in April, it went quiet, then reignited to #3 on the daily board within a day — roughly +1.3k stars in 24 hours — alongside refreshed code, a paper, and Hugging Face weights. Python, Apache-2.0.
Practical angle: A ready-to-experiment open base model for building real-time 3D perception pipelines from low-cost cameras in driver assistance, warehouse robots, or drones.

Tags: #3DReconstruction #GeometricContextTransformer #EmbodiedAI #CameraPoseEstimation #StreamingPerception #Robotics
URL: https://github.com/Robbyant/lingbot-map

KnockOutEZ/wigolo — new daily-board breakout (385 ⭐ total · public beta)

What it does: A local-first MCP server that gives AI coding agents web search, fetch, crawl, and research — with no API keys, no cloud, and $0 per query. Currently in public beta.

How it works: A single request fans out across 18 search-engine adapters, then applies rank fusion plus an ML cross-encoder reranker, attaching an explainable score to each result. URL loading goes through a tiered router (HTTP → TLS impersonation → headless browser) that auto-escalates when it hits anti-bot challenges or SPA shells. It exposes search, fetch, crawl, extract, cache, find-similar, research, and autonomous gather loops through one interface — all running locally under ~/.wigolo/, so nothing leaves the machine.

Use cases & edge: A drop-in replacement for paid search/crawl APIs (e.g. Firecrawl) that lets agents research the web for free and unlimited. Needs Node ≥ 20 and ~1.5 GB of disk; runs on macOS, Linux, and Windows and wires into Claude Code with a single command.
Why it’s rising now: A July 17 post on X spread the pitch that “agents can finally search the web for free,” pushing it onto the daily trending board. Star count is still small (public beta), but its inflow velocity ranks high.
Practical angle: A local web-research backend you can attach to in-house coding agents or RAG pipelines with zero search-API cost — especially fitting where data-exfiltration is a concern.

Tags: #MCP #LocalFirst #WebCrawling #CrossEncoderRerank #HeadlessBrowser #AgentTooling
URL: https://github.com/KnockOutEZ/wigolo

alxndrTL/little-book-rl — daily rank surging #21 → #8 (1.1k ⭐ total)

What it does: A compact learning resource and code repo that walks through reinforcement learning from the basics to applied algorithms — the text runs from Monte Carlo methods all the way to PPO.

How it’s structured: An algos/ folder ships PyTorch implementations of every algorithm covered (MC → … → PPO), and a supplementary/ folder adds detailed explanations and rigorous proofs for the dynamic-programming algorithms only briefly touched in the main text. Theory, math, and runnable code sit side by side in one repo.

Use cases & edge: Positioned as a “little book” — a minimal path through the RL core pipeline you can read and run quickly, instead of a doorstop like Sutton & Barto. V1 shipped in June 2026 under a non-commercial Creative Commons license, with a printed edition on Amazon.
Why it’s rising now: After first trending at daily #21 on July 13, it accelerated into the top 10 within days. Python.
Practical angle: A ready curriculum skeleton for onboarding junior ML engineers or running an internal RL study group, with immediately runnable reference implementations.

Tags: #ReinforcementLearning #PPO #PyTorch #MonteCarlo #PolicyGradient #RLEducation
URL: https://github.com/alxndrTL/little-book-rl

Note: all three repos are first-time entries in this series (not consecutive-day features). Yesterday’s entries (grok-build, Codex-Dream-Skin, and others) were dropped after fresh daily data confirmed their growth had decelerated.