AI Repo: M-Abozaid/esp32-c3-adblock

Skyler: Hey everyone, welcome back to Git News AI — the show where we dig through the hottest trending repositories on GitHub so you don't have to. I'm Skyler, and today we have got a genuinely wild lineup of repos, from a financial markets foundation model with tens of thousands of stars to a plug-and-play memory system for A I agents and a full-blown agentic video production studio. Big thanks to our sponsors keeping this show alive — Ada, which helps teams integrate A I into their actual workflows without the headache, Ago Consulting — that's "Ahh-Go" — your go-to for silicon development consulting all the way from A I to SoC, and Zen Semiconductor, an umbrella company with a whole portfolio of silicon-industry companies, plus they're behind Ada for deploying A I at scale in any business, and their Vect appliance runs local L L Ms and vector search on dedicated hardware which is genuinely cool. Alright, let's get into it. Skyler: Okay so first up we have something a little different but honestly kind of brilliant — it's M-Abozaid's esp32-c3-adblock, and the pitch is exactly what it sounds like: a Pi-hole-class D N S ad blocker running on a two dollar E S P 3 2 dash C 3 microcontroller. Two dollars. That's it. That's the hardware budget. Skyler: So for the uninitiated, Pi-hole is a popular self-hosted network-level ad blocker that most people run on a Raspberry Pi — which is cheap, but it's still like thirty-five to fifty bucks for the hardware. This project squeezes that same idea into a microcontroller that costs about two bucks, and it's actually doing something really clever to make that work. Skyler: The core challenge here is that blocking ads at the D N S level means you need a massive blocklist — we're talking hundreds of thousands of domain names. How do you fit that on a tiny chip with basically no memory? The answer in this repo is elegant: they take five hundred and thirty-seven thousand domains, convert them to forty-bit F N V dash 1 a hashes, and store those in flash memory. Then they do a binary search to look up whether a domain is blocked. Skyler: F N V dash 1 a is a super fast, lightweight hash function — perfect for embedded systems. And binary search means even with over half a million entries, lookups are blazing fast because you're cutting the search space in half with every step. It's algorithmically sound engineering on severely constrained hardware, and that's what makes it fun. Skyler: The repo also includes a U D P D N S sinkhole, which is the mechanism that actually intercepts your D N S requests and either answers them or drops them if the domain is on the blocklist. And on top of that there's a web dashboard so you can actually see what's being blocked on your network. Skyler: The recent activity shows they've got C I set up to compile-check both supported boards on every push and pull request, and they're doing weekly blocklist releases with automated builds. That's a surprisingly mature dev workflow for a hobbyist embedded project. Skyler: The A I relevance here is maybe not direct, but think about it — if you're building edge A I devices, smart home stuff, or any kind of embedded system that needs network filtering, the techniques here around memory-efficient data structures in flash are absolutely transferable. Understanding how to pack and query data on constrained hardware is a real skill for the A I edge computing space. Skyler: This one is sitting at around thirteen hundred and forty-six stars, which is solid for an embedded project, and it's written in C plus plus. If you're into hardware hacking, embedded systems, home networking, or just want a weekend project that'll actually save you from ads for almost nothing, go check out the show notes for the link. Skyler: Next up, and this one is a massive one — fifty-four thousand stars kind of massive — we have Kronos from shiyu-coder, described as a foundation model for the language of financial markets. And yes, financial markets is the domain, and foundation model is not marketing fluff here. Skyler: So what does that mean? A foundation model trained on the language of financial markets means this is a large-scale model pre-trained on financial time series data — price movements, volume, order flow, that kind of thing — the idea being that financial data has its own underlying patterns, its own grammar if you will, and a model trained deeply on it should develop a rich understanding of market behavior that you can then fine-tune for specific tasks. Skyler: This is huge because most people trying to apply A I to finance are either using general-purpose L L Ms and hoping they understand markets, or they're training smaller models from scratch on limited data. Kronos is taking the foundation model paradigm — the same approach that gave us G P T and BERT — and applying it specifically to financial markets. That's a genuinely different bet. Skyler: The Python codebase has been getting active pull request contributions — the recent commits include a batch dimension training fix that came in from a community contributor, which tells me people are actually using this and running into real training issues and solving them together. That's a healthy open source signal. Skyler: For M L engineers in the finance space, this is one to seriously evaluate. If you're working on quantitative trading, risk modeling, portfolio optimization, or really any task that involves understanding market dynamics, having a pre-trained backbone that already speaks the language of markets could dramatically accelerate your work. Skyler: The star count of forty thousand plus puts this in rare company for finance-specific A I repos. It suggests the quant community has taken real notice, which in my experience means either it's doing something genuinely novel or it's producing results that are hard to dismiss. Probably both. Skyler: Honestly, even if you're not in finance, the architectural approach here is worth studying. Applying foundation model thinking to non-text time-series domains is a frontier problem in A I, and seeing a high-quality open-source implementation of it is valuable for any M L engineer thinking about temporal data. Skyler: Check the show notes for the Kronos link — this one could genuinely be a career-relevant tool for anyone in the quantitative finance and M L intersection. Skyler: Alright, repo number three is one I got pretty excited about — it's OptMem by VictorTaelin, and the tagline is: permanent memory for A I agents, a four hundred and twenty-six token prompt, a script, plug and play. Under two thousand stars right now but this feels early and important. Skyler: So here's the problem it's solving. L L Ms have a context window — a finite amount of text they can actually see and reason about at once. For A I agents that need to operate over long periods of time, remember past interactions, and build up knowledge incrementally, this is a real bottleneck. Most memory solutions are complex databases, vector stores, retrieval pipelines — lots of moving parts. Skyler: OptMem's approach is almost provocatively simple. Four hundred and twenty-six tokens is the size of their core memory prompt. That's tiny. The idea is that you give the agent a structured, minimal prompt that defines how it should maintain and update its own memory, combined with a lightweight script, and the whole thing just works without needing a full vector database infrastructure. Skyler: The plug-and-play framing is key here. A lot of agent memory solutions require you to set up Pinecone, Weaviate, or some other vector store, wire up embeddings, build retrieval logic — it's a whole project. If OptMem really delivers on the promise of dropping in a script and a prompt and getting persistent memory, that's a huge reduction in friction. Skyler: The recent activity is interesting — a pull request was merged specifically clarifying that the storage limit is measured in bytes, not characters. That sounds like a tiny detail but it actually matters a lot for developers working with multi-byte character sets or non-English languages. Getting those implementation details right shows attention to correctness. Skyler: The A I relevance score here is high and I think it deserves it. Persistent, efficient memory is one of the core unsolved ergonomic problems in building production A I agents. If you can get that working with just a prompt template and a script, the barrier to building reliable long-running agents drops significantly. Skyler: Who should check this out? Honestly, anyone building with A I agents — whether you're using Claude, G P T, open source L L Ms, whatever — if your agent needs to remember things across sessions without you building a whole memory infrastructure, OptMem is worth an evening of your time. Show notes have the link. Skyler: Number four and this one is enormous — agency-agents by msitarzewski — which is sitting at one hundred and fifty-seven thousand stars. That is not a typo. One hundred and fifty-seven thousand. That puts it in the upper echelon of A I repos on all of GitHub. Skyler: The description calls it a complete A I agency at your fingertips, and what this repo actually is is a massive collection of specialized A I agent definitions — each one is described as having personality, processes, and proven deliverables. We're talking frontend wizards, Reddit community agents, whimsy injectors — yes, that's a real agent in there — and reality checkers. Skyler: So what's the actual structure here? These are agent definition files — think of them as detailed system prompts and behavioral specifications that turn an L L M into a focused expert with a specific role, a specific personality, and a specific workflow. It's written primarily in Shell, which suggests a lot of the tooling around deploying and managing these agents is script-based and designed to be composable. Skyler: The creative naming here isn't just for fun — the specificity of each agent's role is the whole point. Instead of prompting a generic A I assistant, you're invoking a specialized agent that has been carefully designed to behave consistently within its domain. That's a real pattern for building reliable A I workflows. Skyler: The star count suggests this has gone quite viral, and the recent activity confirms it — they're merging batches of nearly fifty pull requests at a time from community contributors fixing and expanding content. That's a repo with serious community momentum. Skyler: From a developer perspective, the interesting question is: what can you build by composing these agents together? An actual digital agency workflow where a frontend wizard hands off to a copywriter who gets checked by a reality checker — that multi-agent orchestration pattern is where production A I systems are heading. Skyler: Even if you don't use the agents directly, studying how each one is defined — the personality framing, the process steps, the deliverable structure — is genuinely instructive for anyone building their own agent prompts. There's a craft to writing good agent definitions and this repo has apparently crowdsourced a lot of it. Skyler: For A I developers, product teams, content agencies, or anyone thinking about how to operationalize L L Ms for specific roles, agency-agents is worth digging into. It's essentially an open-source library of agent blueprints, and with that star count, it's clearly resonating with a huge audience. Show notes have the link. Skyler: And our fifth and final repo today is OpenMontage by calesthio, which describes itself as the world's first open-source agentic video production system. Sixty-four thousand stars, Python, and the scope of this thing is genuinely wild. Skyler: Let's break down what agentic video production actually means. Traditional video production involves a pipeline of tasks — scripting, asset generation, scene composition, editing, audio, rendering — usually handled by different tools and different people. OpenMontage is attempting to orchestrate all of that using A I agents, and the numbers here are staggering: twelve production pipelines, over a hundred tools, and more than seven hundred agent skill and production knowledge files. Skyler: Seven hundred. That's not a typo either. Seven hundred specialized knowledge and skill files that teach the system how to do things like manage temporal latency in video generation — which is actually referenced in the recent pull requests — handle different video providers, manage scene transitions, all the craft knowledge of video production encoded for agents. Skyler: The pitch to developers is particularly clever: turn your A I coding assistant into a full video production studio. So if you're already using Cursor, Windsurf, Claude Code, or any A I coding tool, the idea is that OpenMontage gives that assistant the production knowledge and tooling to actually produce videos end-to-end. That's a very practical integration angle. Skyler: The recent commits show active work on video model integrations — there's a fix for WAN 2.2 text-to-video temporal latent handling, which is a reference to one of the newer open video generation models. This tells me the project is keeping pace with the fast-moving video A I ecosystem and integrating new models as they emerge. Skyler: For content creators, marketers, A I researchers working in the video space, or developers who want to explore what a full agentic production system looks like under the hood, OpenMontage is a fascinating case study. The architecture alone — how you decompose video production into agent skills and chain them across pipelines — is worth studying even if you never produce a single video with it. Skyler: The sixty-four thousand star count and the active contributor community suggest this one has real legs. Agentic video is still early but it's clearly capturing imaginations, and an open-source system at this scale gives the community a foundation to build on. Link is in the show notes — this one is worth a deep dive. Skyler: And that's a wrap on today's episode of Git News AI — five repos ranging from a two-dollar hardware ad blocker to a hundred-and-fifty-seven-thousand-star agent library, and everything in between. Massive thank you to our sponsors: Ada for helping teams actually integrate A I into their work, Ago Consulting — "Ahh-Go" — for silicon development expertise from A I all the way to SoC, and Zen Semiconductor for their portfolio of silicon-industry companies and for the Vect appliance that brings local L L Ms and vector search to dedicated hardware. These folks make this show possible. Skyler: If you want more A I news and deep dives, check out our sister show AI Special Report on the Ada podcast network — it's a great companion to what we do here. And if you're not subscribed yet, hit that subscribe button wherever you're listening — new episodes drop regularly and we're always finding the repos you need to know about. Skyler: All the links for today's repos are in the show notes — go explore, go build, and I'll catch you in the next one.