GitHub Trending Weekly #46: latticedb, Gradient, buzz, learn, workout-guide, rome, CarWatch, selfdb Transcript, AI Summary & Key Points

00:00 Welcome back to GitHub awesome. This is GitHub trending weekly number 46. 35 trending open source projects on GitHub right now. Let's go. Lattis DB is an embedded graph database in a single file written in Zigg with no dependencies that does relationship traversal, vector similarity, and full text ranking in one query language over the same data. So a rag query can match a chunk by embedding hop to its document then to the author in one cipher statement rather than gluing three systems together.

00:32 One file, one writer, no server. Gradient trains research agents with reinforcement learning while leaving the whole loop open for inspection. An agent searches a synthetic company workspace, cites evidence, and records every tool call. gradient scores correctness, citation quality, and efficiency, then uses openpipe art with GRPO to train an adapter and compare it with the base model on held out tasks.

01:01 Buzz is a self-hosted team workspace where humans and AI agents join the same rooms with separate identities and audit trails. Messages, reactions, approvals, workflows, and Git activity become signed Nostra events in one searchable log. So an agent can discuss a bug, submit a patch, and leave the evidence beside the decision. Channels, desktop clients, canvases, search, git events, and YAML workflows work today.

01:27 Wallgit puts a git server in one binary backed by S3 or GCS with no database or elected leader. The bucket is the repository. Each server's disk is just a disposable cache. Pushes become immutable log entries, while a compare and swap manifest decides which update wins, letting multiple instances share storage safely. It includes smart HTTP policies, web hooks, and a web UI.

01:52 Ambient context reads the text of whatever window you're focused on through the Mac OS accessibility tree, and appends it to a markdown file, one per day, in a folder you choose. No screenshots, no video, no network calls. Password managers and private browsing are excluded outright. Secure fields are skipped at the source and the capture folder excludes itself so it can't observe its own output.

02:17 Learn is the author's own PI configuration from his video on learning with AI published as is. The repo is api directory. You clone it as your project's config. A teach skill holds the philosophy. A visualized skill adds a diagram only when a picture is clearer and sub agents split off for research and for making SVG and mermaid. He says it's written for one learner, him.

02:38 No license file. Workout guide packages. 302 exercise illustrations for apps that need consistent movement visuals without commissioning every pose. Each exercise includes three matching frames, searchable metadata, typed lookup helpers, and transparent SVG assets that work without tying you to React or another framework. There's also a static gallery for browsing the catalog.

03:07 Rome's bet is that the environment matters more than the model, so its unit isn't a chat, but an app, a manifest, plus typed actions, appowned agents, skills, hooks, a web UI, and its own database that survives the conversation. Describe something missing and Rome scaffolds it into your instance as ordinary git tracked source rather than hidden model state.

03:28 Docker quickart binds to loop back and telemetry export stays off unless you point it somewhere. Dupe is a multiplayer design canvas where humans and AI agents edit the same work live. Each frame renders HTML inside a sandboxed iframe while cursors, comments, agent status, and changes stream through a websocket room. Its built-in MCP server lets agents create, revise, screenshot, and review frames under the user's access.

03:56 You can self-host it with Docker. OCR. It is a browser extension that pins one screen region, then OCRs that exact rectangle every time you hit the hotkey, appending to a running transcript. Tessoract is bundled, so it runs offline with no outbound requests at all. The permissions model is the unusual part. No site access at install. Single captures ride on active tab.

04:19 And only an auto run or a cross origin iframe needs a durable grant you approve per site. Caratch puts a Raspberry Pi 5 in the car running a 35 billion parameter model locally at about 3.5 tokens a second, answering from the car's own 745page owner's manual with page citations and refusing what the manual doesn't cover. Voice runs on PI, no wake word.

04:43 What's rare is the status table. Proven means it worked on the real car and the OBD reader is marked unverified against the actual vehicle. Better voice is Mac OS dictation that takes the screen with it. Hold option and talk. And while you're recording, circle anything on screen with the pointer. A blue trail follows and each circle captures the display.

05:08 The transcript lands in your text field with those screenshots after it. Transcription runs locally on parakeet with an opt-in 36 megabyte grammar model that falls back to the raw transcript if it fails. Self DB makes a program that is a SQLite database instead of an ELF file. Its Linux interpreter rebuilds or maps that data at launch while ordinary SQL can inspect dependencies, remove sections, or update embedded content.

05:36 The included web server takes the idea literally, keeping code, routes, pages, and visitor logs in one file. It's an experimental Nixxosheavy format with multiple loader modes, not a drop-in replacement for normal binaries. Proliferate runs clawed codecs through their own native harnesses rather than reimplementing them. Each task getting its own git work tree, so the branch, terminal, conversation, and review state isolated.

06:04 Agents delegate to sub agents and collect the results. MCPs, skills, and custom tools are configured once and shared by every agent. System Atlas turns architecture discussions into two synchronized views. An interactive isometric map for people and a generated system. MD for the repository. Both come from one data file, so structures, flows, decisions, and open questions change together.

06:30 Chapters reveal the system gradually while clickable data packets expose roots and sample payloads. Karna starts a real OCI container in about 3.5 milliseconds from a 1.52 megabyte static binary rootless with no Damon and nothing running at rest. Its whole Rust dependency tree is lib C. JSON and OCI manifests are parsed by hand and pull shells out to the curl and tar already on the machine.

06:55 Resource profiles for CPU, memory, disk, and devices are declared once in a current TOML and attached by name. IDO searches files by what you remember, not which machine or path you stored them on. Its current Windows first build keeps a readonly catalog of NTFS and SMB sources, indexes names and file contents, and preserves stale results when a machine goes offline.

07:20 Every answer carries provenence and source completeness. So a missing result can admit what wasn't checked. Agent trail shows what a coding agent is doing now, not just the transcript it leaves behind. It watches plans, tool calls, and file changes, then maps that activity onto repository components with working, blocked, and completed states. Claude Code gets richer run cards through local hooks.

07:46 Codec, cursor, and other agents still work through file watching. Backpass treats agents.mmd like a model you improve from its own mistakes. It finds local coding agent transcripts, strips tool noise and obvious secrets, then looks for repeated failures that justify a small instruction change. Every proposal includes quotes from real sessions. New rules need evidence from at least two sessions, and nothing writes until you approve it.

08:16 Rataflow brings React Flow style node editors into a Rust terminal interface. Nodes and edges can carry your own types while the widget handles automatic layout, dragging, connections, resizing, selection, pan and zoom, keyboard navigation, themes, and savable graph snapshots. You still own the event loop and every interaction returns semantic events instead of mutating hidden state.

08:41 Centio gives each AI agent a real mailbox without stitching together IMAP polling, message parsing, SMTP, and web hooks yourself. Incoming mail arrives as structured authenticated events while agents reply in thread the rest or a small MCP server. Tenency reaches domains, keys, rate limits, suppressions, and reputation. So one deployment can isolate many customers.

09:04 Vault graph turns an obsidian vault into a deterministic disc instead of a forced directed hairball. Top level folders become wedges. Linked notes move toward the center and a timeline lets you replay the vault's growth without rearranging nodes unpredictably. It works as an Obsidian plugin or exports an offline HTML file with no server or network calls.

09:28 Omari Pods gives AirPods a Linux control panel that understands the hardware. It shows battery levels for both buds in the case, exposes only the listening modes each model supports, and adds adaptive noise, conversation awareness, onebud ANC, and ear detection. A bundled Libre Pods Demon handles Apple's protocol and publishes state to a file, so the panel itself doesn't pull Bluetooth.

09:52 Own me keeps coding agent memory where the code lives. Reviewed markdown inside the repository. Claude code, codeex, cursor, and grock can recall the same decisions in debugging lessons, while a deterministic BM25F index ranks them locally without a model call or network request. Git handles sharing, review, and roll back and audits limit duplicates and bloat.

10:17 Factory installs a repeatable software delivery loop into a GitHub repository without adding a custom orchestrator. issues become the cue. Fresh clawed code or codec sessions triage bounded work implemented on deterministic branches, run project gates, and hand the diff to an independent verifier before opening a draft pull request. The charter defines what agents may touch, and humans still own every merge.

10:44 Monode puts several coding agent CLIs behind one desktop interface without becoming another model provider. If Claude code, codeex, cursor or FX is already installed and authenticated, you can open sessions as tabs and send prompts from the same composer while each tool keeps its own account. Source builds support Mac OS and Linux, but the downloadable release is Apple silicon only.

11:09 Kimoto.cpp brings Nvidia's Komodo texttootion model into a native C++ and GGML runtime. Give it a UTF8 prompt or a precomputed embedding and it generates SNPLX22 rotations and root movement locally through CPU or Vulcan. The port includes GGF loading, weight conversion, DDM sampling, C and C++ APIs, parody tests, and a browser demo. Frontier agent packages longunning research and file work into a terminal agent you can actually inspect.

11:36 React keeps one stateful worker on the task while agent team gives a coordinator a live taskboard and bounded parallel sub aents. Inputs stay red only. Writes show diffs for approval. Every action is traced and sessions can resume or revert. AI engineering lab is a free self-paced 24-week course that takes you from Python basics to deployed AI systems.

12:02 Instead of disconnected demos, every week extends the same fictional freight company, turning its data into models, retrieval systems, agents, and cloud pipelines. Runnable notebooks accompany each topic. And from week three, every artifact needs a metric and error note. Open sheet lets coding agents build spreadsheets as React models instead of juggling fragile cell addresses.

12:24 You name fields like revenue or tax rate and the framework resolves their final Excel coordinates after rows and layouts settle. Its formula tree powers the browser preview and exports live formulas to XLSX so recipients can change assumptions and recalculate normally. CSV, HTML, and PDF exports are included too. Try Omari runs DHH's Omari desktop, the Arch and Hyperland setup as a native app on an Apple silicon Mac with nothing to configure.

12:55 It packages a pinned ARM 64 Arch image, a QUU runtime on Apple's hypervisor framework, and a Swift launcher into one signed app with hardware accelerated VGL graphics, a resizable window that resets guest resolution live and Mac audio routed in. Halo puts one governance boundary between company knowledge and the agents using it. Credentials resolve identity, role, namespace, and source scope on the server.

13:22 So agents can't simply claim broader access in an MCP request. Reads and writes carry provenence, policy decisions, and appendon audit events, while deletion produces a signed eraser certificate. Speculative programmatic tool calling speeds up agents that write code calling other models. Normally the harness finishes generating a whole code block then runs it and only then fires the slow sublm calls inside one after another.

13:50 This library reads the code as it streams and launches those calls early as futures so they run during generation instead of after. Elemental sandbox is a 3JS VFX workbench for designing game abilities while they're still moving. Seven casts share five targeting shapes, then build their effects from procedural geometry, handwritten GLSL, and GPU particles instead of sprite sheets.

14:15 Press pause mid eruption, and every slider remains live, so you can reshape timing, color, particles, and silhouettes against a frozen frame. Haimeimi does realtime speechtoext on a CPU, no GPU, and nothing sent to the cloud. And its trick is refusing to use one model for everything. It detects the language of each utterance in the first few seconds, then routes it to a specialist, a Japanese model, a Chinese one, an English one, with a 1,600 language fallback behind them, all quantized and running through Sherpa Onx.